{"success":true,"data":{"username":"yena","posts":[{"id":"2072765754102174114","shareId":"bdzCi1AGuM","platform":"x","url":"https://x.com/aakashgupta/status/2072765754102174114","title":"Bridgewater가 방금 발표한 수치들은 모든 프론티어 연구소를 긴장하게 만들 만한 것들입니다.","shareMode":"full","createdAt":1783052321711,"postedAt":"2026-07-02T19:34:00.000Z","publishedDate":"2026-07-02T19:34:00.000Z","archivedAt":"2026-07-03T04:18:38.426Z","archivedDate":"2026-07-03T04:18:38.426Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/bdzCi1AGuM","contentType":"article","content":{"text":"# Bridgewater가 방금 발표한 수치들은 모든 프론티어 연구소를 긴장하게 만들 만한 것들입니다.\n\n## 1\n\nBridgewater가 방금 발표한 수치들은 모든 프론티어 연구소를 긴장하게 만들 만한 것들입니다.\n\n세계 최대 헤지펀드가 투자자들이 매일 하는 여섯 가지 문서 필터링 작업에서 Gemini, Claude, GPT를 테스트했습니다. 단순한 프롬프트는 약 50%의 점수를 받았습니다. 동전 던지기 수준이죠. 전문가가 작성한 프롬프트는 정확도를 78%까지 끌어올렸습니다. 투자자들은 워크플로에 시스템을 신뢰하려면 80%가 필요하다고 했는데, 어떤 프론티어 모델도 이를 넘지 못했습니다. GPT 5.4는 5.2보다 43% 더 비쌌고, 정확도도 겨우 조금 나았을 뿐입니다.\n\n그래서 그들은 Tinker에서 Qwen3-235B를 파인튜닝했습니다. 정확도 84.7%. 최고의 프론티어 모델보다 실수가 29.8% 적었습니다. 추론 비용은 1/14 수준으로요.\n\n가장 영리한 부분은 논문 중간에 묻혀 있습니다. 그들의 벤더가 라벨링한 훈련 데이터는 잘못된 라벨로 가득 차 있었고, 전문가 라벨링 비용은 모든 데이터에 적용하기엔 너무 비쌌습니다. 그들의 해결책: 노이즈가 섞인 데이터셋으로 모델을 훈련시킨 뒤, 그 모델을 다시 훈련 데이터에 돌려보는 거였습니다. 모델이 동의하지 않는 예시는 모두 선임 투자자들에게 라우팅됐는데, 그 이유는 예시가 진짜 어려웠거나 라벨이 틀렸기 때문입니다. 모델의 혼란이 나쁜 라벨의 탐지기가 된 셈이죠.\n\n프롬프팅은 구조적인 이유로 한계에 부딪혔습니다. 프롬프트는 전문가가 말로 표현할 수 있는 판단만 담을 수 있습니다. 중앙은행 메모 중 실제 금리 변동을 시사하는 것에 대한 20년의 직관은 지시어로 압축되지 않습니다. 그것은 라벨링된 예시를 통해 전달됩니다.\n\n수십 년간의 전문가 결정 데이터를 보유한 모든 기관이 이제 그 아카이브가 프론티어를 그들만의 특정 업무에서 이기는 모델을 훈련시킬 수 있다는 걸 알게 됐습니다. 알파는ずっと 서랍장 안에 있었네요.\n\n---\n\n## 2\n\nTo get all my takes without an algorithmic filter, subscribe to my newsletter:\n\nhttps://t.co/vHdRoFmd5z"},"author":{"name":"Aakash Gupta","handle":"aakashgupta","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/146c61077c95cdda/avatar-d473e8da2a65.jpg","bio":"✍️ https://t.co/8fvSCtBv5Q 💼 https://t.co/STzr4nqxnm 🤝 https://t.co/SqC3jTyP03 🎙️ https://t.co/fmB6Zf5UZv"},"metadata":{"timestamp":"2026-07-02T19:34:00.000Z"},"media":[],"comments":[{"id":"2072882938773901701","content":"You wouldn't want a random person work on your electrical panel in your home. You wouldn't want an untrained model working your your repo either.","author":{"name":"Tim","username":"MrOcelot1976","handle":"MrOcelot1976","url":"https://x.com/MrOcelot1976"},"timestamp":"2026-07-03T03:19:39.000Z"},{"id":"2072820062101311862","content":"Not really surprising, doubly so if you have done any work for the labelling companies that subcontract to the frontier labs.","author":{"name":"The Bayesian Bushman","username":"JackarooFawkes","handle":"JackarooFawkes","url":"https://x.com/JackarooFawkes"},"timestamp":"2026-07-02T23:09:48.000Z"},{"id":"2072796237158592954","content":"yeah they just fucking suck at prompting and harnessing you do not need to fine tune a model for doc scanning lmfao\n\nit’s like asking a bunch of horae carriage owners to test your new gas motor and they use it to power an automated horse whipper n complain abt its efficiency LOL","author":{"name":"Bob O","username":"BobO58076053213","handle":"BobO58076053213","url":"https://x.com/BobO58076053213"},"timestamp":"2026-07-02T21:35:08.000Z"},{"id":"2072813032862831014","content":"This post is literally written by AI about I and on first read does not articulate very well what it’s trying to say. I’m a paid subscriber so I’m hoping there’s a far better explanation on your substack…","author":{"name":"GP","username":"eiregp","handle":"eiregp","url":"https://x.com/eiregp"},"timestamp":"2026-07-02T22:41:52.000Z"},{"id":"2072782319476261015","content":"Or you could just ask the old guy in the corner","author":{"name":"HundredDollarBillz","username":"Jhawkbill","handle":"Jhawkbill","url":"https://x.com/Jhawkbill"},"timestamp":"2026-07-02T20:39:50.000Z"},{"id":"2072875538654368193","content":"Maybe we can use AI to decipher WTF this AI post is trying to say.","author":{"name":"Londonman","username":"Londonman90","handle":"Londonman90","url":"https://x.com/Londonman90"},"timestamp":"2026-07-03T02:50:15.000Z"},{"id":"2072862922796081567","content":"Garbage take. The eval was broken at 50% performance on a binomial task by frontier models. The real movement was 77 to 78 to 85. The paper says this if you parse the claims.","author":{"name":"Henry Walton Jones Jr.","username":"fugoaway","handle":"fugoaway","url":"https://x.com/fugoaway"},"timestamp":"2026-07-03T02:00:07.000Z"},{"id":"2072820780115828786","content":"defining the task precisely enough to fine-tune is the actual unlock. most teams default to frontier because they skip that step.","author":{"name":"Alexander Benz","username":"alexanderbenz","handle":"alexanderbenz","url":"https://x.com/alexanderbenz"},"timestamp":"2026-07-02T23:12:39.000Z"},{"id":"2072826852947746871","content":"That why you need PLTR","author":{"name":"p_tom40","username":"p_tom40","handle":"p_tom40","url":"https://x.com/p_tom40"},"timestamp":"2026-07-02T23:36:47.000Z"},{"id":"2072824222683111752","content":"How did they handle the initial noisy labels?","author":{"name":"Steven Cheng","username":"stevencheng","handle":"stevencheng","url":"https://x.com/stevencheng"},"timestamp":"2026-07-02T23:26:20.000Z"},{"id":"2072798106501222525","content":"This should not be a surprise. All the benchmarks around the last human test or whatever random puzzle test is useless for professionals and enterprises that have a very nuanced job.","author":{"name":"Mike Lin","username":"mikelin1789","handle":"mikelin1789","url":"https://x.com/mikelin1789"},"timestamp":"2026-07-02T21:42:33.000Z"},{"id":"2072853084280131654","content":"fine-tuning helps, but the hard part is keeping labels stable when analysts disagree","author":{"name":"Sebastian Buzdugan","username":"sebuzdugan","handle":"sebuzdugan","url":"https://x.com/sebuzdugan"},"timestamp":"2026-07-03T01:21:01.000Z"},{"id":"2072834566075633708","content":"meta Data is important ,\n@elonmusk\n\n@xai\n as demonstrated in results .","author":{"name":"Rick ;-) Classroom MBA 5D viewpoint 40 yrs in Mkts","username":"REBIII333Re","handle":"REBIII333Re","url":"https://x.com/REBIII333Re"},"timestamp":"2026-07-03T00:07:26.000Z"},{"id":"2072809621874503757","content":"I'd like to run this premise by you.\n\nData, or availability of data, is plentiful.\n\nAn AI model can be trained and retrained ad infinitum to generate labels, that's not the scarce resource either.\n\nThe bottleneck to scaling is moving toward the costs of re-deriving the","author":{"name":"Stephanie So","username":"ComplicatedIsOK","handle":"ComplicatedIsOK","url":"https://x.com/ComplicatedIsOK"},"timestamp":"2026-07-02T22:28:19.000Z"},{"id":"2072810292334260716","content":"Train the models with Ray Dalio's Principles","author":{"name":"Mark Nakata","username":"mrnonel","handle":"mrnonel","url":"https://x.com/mrnonel"},"timestamp":"2026-07-02T22:30:59.000Z"},{"id":"2072790763352490212","content":"This proves that hyper-targeted local weights completely erase the premium value of generalized frontier endpoints.","author":{"name":"Ilman Shazhaev","username":"shzhv13","handle":"shzhv13","url":"https://x.com/shzhv13"},"timestamp":"2026-07-02T21:13:23.000Z"},{"id":"2072798453273391449","content":"So what you are saying is that current frontier models with good prompting and no fine tuning are sufficiently accurate for investors to trust the system in their workflows? Idk why this should make frontier labs nervous.","author":{"name":"joncelery","username":"johncelery","handle":"johncelery","url":"https://x.com/johncelery"},"timestamp":"2026-07-02T21:43:56.000Z"},{"id":"2072798049064407121","content":"Bridgewater’s success rate since its inception has been less than 50% so the models are definitely doing better","author":{"name":"AKumar","username":"parwalkibarfi","handle":"parwalkibarfi","url":"https://x.com/parwalkibarfi"},"timestamp":"2026-07-02T21:42:20.000Z"},{"id":"2072799676676399420","content":"What might the consequences to the users be for the 20% inaccuracy which is tolerated?\n\nWhat types of workflows can tolerate a 20% error rate?","author":{"name":"Cool Master","username":"CoolMaster312","handle":"CoolMaster312","url":"https://x.com/CoolMaster312"},"timestamp":"2026-07-02T21:48:48.000Z"},{"id":"2072802064678228208","content":"1/14th the cost....","author":{"name":"Ab","username":"Ab83126691","handle":"Ab83126691","url":"https://x.com/Ab83126691"},"timestamp":"2026-07-02T21:58:17.000Z"}],"aiComments":[]},{"id":"2072676216507887670","shareId":"QYwsEAD89I","platform":"x","url":"https://x.com/acidsound/status/2072676216507887670","shareMode":"full","createdAt":1783052283753,"postedAt":"2026-07-02T13:38:13.000Z","publishedDate":"2026-07-02T13:38:13.000Z","archivedAt":"2026-07-03T04:18:02.150Z","archivedDate":"2026-07-03T04:18:02.150Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/QYwsEAD89I","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/ddhYiZ9ymF/media/0.png","contentType":"post","content":{"text":"https://github.com/googlecolab/google-colab-cli\n\n너무 쩌는데? 역시 대감집 인심이 최고다.\n\n[GitHub - googlecolab/google-colab-cli](https://t.co/wW0QnBzx11)"},"author":{"name":"이재호","handle":"acidsound","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/0dbc36da0e63223e/avatar-ad13d7fb0eae.jpg"},"metadata":{"likes":170,"shares":43,"views":11223,"timestamp":"2026-07-02T13:38:13.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ddhYiZ9ymF/media/0.png","type":"image","width":1200,"height":600}],"aiComments":[]},{"id":"2072746707285147994","shareId":"7MybyWg7cC","platform":"x","url":"https://x.com/aresotik/status/2072746707285147994","shareMode":"full","createdAt":1783052270512,"postedAt":"2026-07-02T18:18:19.000Z","publishedDate":"2026-07-02T18:18:19.000Z","archivedAt":"2026-07-03T04:17:48.634Z","archivedDate":"2026-07-03T04:17:48.634Z","viewCount":3,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/7MybyWg7cC","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/znV9SPFNrk/media/thumb-0.jpg","contentType":"post","content":{"text":"esta herramienta analiza cualquier web y te genera un sistema de diseño completo listo para replicar.\n\nle pegas una URL y te devuelve estructura, estilos y layouts organizados desde cero.\n\nse llama Aura Build, y es gratis. \n\nguárdala 🔖 https://t.co/EGHFFiZD2H"},"author":{"name":"ares. 🎧","handle":"aresotik","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/075ea155a1c3b7c0/avatar-1abf1536e1ac.jpg","bio":"26 | making AI easy to learn | @aiscwork partner ✦ aresotik1@gmail.com"},"metadata":{"likes":227,"shares":68,"views":6354,"timestamp":"2026-07-02T18:18:19.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/znV9SPFNrk/media/0.mp4","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/znV9SPFNrk/media/thumb-0.jpg","type":"video","width":2880,"height":2160,"duration":50}],"aiComments":[]},{"id":"2072696009244004779","shareId":"CWNLLzDJJt","platform":"x","url":"https://x.com/shotovim/status/2072696009244004779","shareMode":"full","createdAt":1783052222803,"postedAt":"2026-07-02T14:56:52.000Z","publishedDate":"2026-07-02T14:56:52.000Z","archivedAt":"2026-07-03T04:16:59.602Z","archivedDate":"2026-07-03T04:16:59.602Z","viewCount":3,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/CWNLLzDJJt","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/sOLg2URXIC/media/0.jpg","contentType":"post","content":{"text":"AIの長期記憶や会話の保存、Obsidianのvaultに置くより独自のシステムを組んだほうがAI親和性が高い。\n\nGizmodoの動画でも言われたけどvaultは人間向けのノートで、AIが毎回読むにはノイズが多いから。\n\n親和性高めるためにフォルダ構成とフロントマターを意識して設計するのがよさそう。\n\nいまわたしが構築している記憶システムはこんな感じ。"},"author":{"name":"松濤Vimmer","handle":"shotovim","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/c8ce7e552764280f/avatar-3ce8fa9f66e7.jpg","bio":"CyberAgent | SWE | Obsidian × AI、効率化、技術に関する情報を発信 | Vim、Cursor、Obsidian、Raycast、Arc を好みます | 投稿は一個人の意見です | お仕事のご依頼はフォームまたはDMまで"},"metadata":{"likes":347,"shares":23,"views":42826,"timestamp":"2026-07-02T14:56:52.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/sOLg2URXIC/media/0.jpg","type":"image","width":2408,"height":1358}],"quotedPost":{"id":"2072316981622268341","platform":"x","url":"https://x.com/shotovim/status/2072316981622268341","shareMode":"full","content":{"text":"これ激しく同意でAIの長期記憶用のみで使うみたいな活用だったらObsidianじゃなくていいんだよね。\n\nファイルを参照する・開くアプリの回数が ClaudeCode・Codex >>> Obsidian みたいなひとはObsidian Vaultにするのではなく最適化された記憶管理システムを構築するのがいい。 https://t.co/i75oRcvYHx"},"author":{"name":"松濤Vimmer","handle":"shotovim","avatar":"https://pbs.twimg.com/profile_images/1846880148551045120/IkLrswuq_200x200.jpg"},"metadata":{"likes":158,"shares":8,"timestamp":"2026-07-01T13:50:44.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/sOLg2URXIC/quoted/0.jpg","type":"image","width":1199,"height":677}]},"aiComments":[]},{"id":"2072528567049556204","shareId":"lFkmyo6YGD","platform":"x","url":"https://x.com/markletree/status/2072528567049556204","shareMode":"full","createdAt":1783052147401,"postedAt":"2026-07-02T03:51:30.000Z","publishedDate":"2026-07-02T03:51:30.000Z","archivedAt":"2026-07-03T04:15:41.355Z","archivedDate":"2026-07-03T04:15:41.355Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/lFkmyo6YGD","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/i12r3LYOwS/media/0.jpg","contentType":"post","content":{"text":"Companies are scaling to thousands of custom agents that call tools and each other. But most of those calls happen directly, with no verified identity and no authorization check.\n\nInstead, each agent should have its own identity, a short-lived, attested credential rather than a static API key. When it's acting for a user or another agent, its identity should say so.\n\nAn agent gateway enforces this. Every call routes through the gateway, and before it reaches a tool or another agent, that endpoint verifies the agent's identity, uses it to authorize which tools (MCPs, APIs, CLIs) and other agents (A2A) the agent can reach, and audits the call. Every action then traces back to the user who triggered it, and no agent can exceed the permissions of whoever it's acting for.\n\nAnd there are already strong open source and vendor options, so you don't need to build this yourself."},"author":{"name":"mark","handle":"markletree","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/a5e8876f2d0d3157/avatar-2fbd1204eda4.jpg"},"metadata":{"likes":52,"shares":10,"views":4334,"timestamp":"2026-07-02T03:51:30.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/i12r3LYOwS/media/0.jpg","type":"image","width":1200,"height":675}],"aiComments":[]},{"id":"2072637407762202648","shareId":"Uhtq09CwUi","platform":"x","url":"https://x.com/cevenif/status/2072637407762202648","shareMode":"full","createdAt":1783051965556,"postedAt":"2026-07-02T11:04:00.000Z","publishedDate":"2026-07-02T11:04:00.000Z","archivedAt":"2026-07-03T04:12:44.128Z","archivedDate":"2026-07-03T04:12:44.128Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/Uhtq09CwUi","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/ZzkIuWFHsd/media/0.png","contentType":"post","content":{"text":"这个 chrome-devtools-mcp 项目，如果你还没关注，那损失可真不小。\n\n它做了一件相当硬的事——把 Chrome DevTools 的完整调试能力，直接开放给了 AI 编程助手。\n\n常遇到的烦恼：\n\n1️⃣ 修改代码、刷新页面、查看报错。\n\n2️⃣ 切换到 DevTools 面板，手动检查 DOM 树、过滤网络请求、分析控制台输出。\n\n3️⃣ 回到编辑器修改代码，再重复上述步骤。\n\n4️⃣ 在编辑器与浏览器之间反复横跳，不知不觉半天就过去了。\n\nAI 编程助手自己就能直接操控 DevTools 进行深度检查。无论是复杂的 DOM 结构、纷繁的网络请求，还是性能追踪记录，它都能自行查看、研判并尝试修复。\n\n整个流程中最消耗时间、最令人头疼的反复试错与猜测环节，被这个工具彻底砍掉了。\n\n前端联调中的协作沟通、线上问题的回归排查、页面加载的性能诊断，这最折磨人的场景，现在都有了针对性的解决方案。\n\n🔗 : https://github.com/ChromeDevTools/chrome-devtools-mcp"},"author":{"name":"Gas留在链上","handle":"cevenif","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/7c675bae26786dfe/avatar-984c5ae9d0f7.jpg"},"metadata":{"likes":11,"shares":4,"views":1802,"timestamp":"2026-07-02T11:04:00.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ZzkIuWFHsd/media/0.png","type":"image","width":877,"height":730}],"aiComments":[]},{"id":"2072753169751490814","shareId":"JoOPJ9dUfi","platform":"x","url":"https://x.com/Zephyr_hg/status/2072753169751490814","shareMode":"full","createdAt":1783051949315,"postedAt":"2026-07-02T18:44:00.000Z","publishedDate":"2026-07-02T18:44:00.000Z","archivedAt":"2026-07-03T04:12:26.883Z","archivedDate":"2026-07-03T04:12:26.883Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/JoOPJ9dUfi","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072692963621122048/img/ytaFca_mjBsl1Hlc.jpg","contentType":"post","content":{"text":"Sam Altman, CEO of OpenAI: \n\n\"To build with AI, you don't need a huge team. You need a good idea.\"\n\nIn a 52-minute keynote, he describes how building stopped requiring a company behind you.\n\nFor the first time, one person with a setup can produce what used to take a floor of staff.\n\nThat's the shift making the first people rich alone, without a single hire.\n\nWatch the keynote, then see the one thing they learned in the article below.\n\nBookmark this one."},"author":{"name":"Zephyr","handle":"Zephyr_hg","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/95984c088d01186b/avatar-e73c0e90cda3.jpg","bio":"I teach solo professionals to build AI systems that do a team's work. Free systems weekly. 12,000+ already winning with AI ↓"},"metadata":{"likes":118,"shares":38,"views":17821,"timestamp":"2026-07-02T18:44:00.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2072692963621122048/vid/avc1/1280x720/TN5bBP5IgF4Gv1t2.mp4?tag=28","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072692963621122048/img/ytaFca_mjBsl1Hlc.jpg","type":"video","width":1280,"height":720,"duration":3129}],"quotedPost":{"id":"2072692882000011353","platform":"x","url":"https://x.com/Zephyr_hg/status/2072692882000011353","shareMode":"full","content":{"text":"The First People Getting Rich Alone With AI All Learned One Thing\n\nhttp://x.com/i/article/2072691108904099840"},"author":{"name":"Zephyr","handle":"Zephyr_hg","avatar":"https://pbs.twimg.com/profile_images/1741076613939277824/vs_v9C1j_200x200.jpg"},"metadata":{"likes":63,"shares":8,"timestamp":"2026-07-02T14:44:26.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kIdtUdHGXc/quoted/0.jpg","type":"image","width":1200,"height":480}]},"aiComments":[]},{"id":"2072659051134218470","shareId":"TO1mLqibql","platform":"x","url":"https://x.com/pauliusztin_/status/2072659051134218470","shareMode":"full","createdAt":1783051777652,"postedAt":"2026-07-02T12:30:00.000Z","publishedDate":"2026-07-02T12:30:00.000Z","archivedAt":"2026-07-03T04:09:36.111Z","archivedDate":"2026-07-03T04:09:36.111Z","viewCount":7,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/TO1mLqibql","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/yIjVEyhRne/media/thumb-0.jpg","contentType":"post","content":{"text":"I found one of the most comprehensive AI engineering guides ever.\n\nAnd the majority of the resources are FREE!\n\n→ https://github.com/louisfb01/start-ai-engineering\n\nMost people trying to learn AI engineering end up doing the same thing:\n\nWatching random YouTube videos\n\nFollowing every new framework\n\nJumping between tutorials\n\nNever building enough\n\nThis guide takes a much more structured approach.\n\nInstead of organizing resources by popularity...\n\nIt organizes them around the skills you need.\n\nFor example:\n\nContext engineering\n\nRAG\n\nMCP\n\nAgent architectures\n\nEvals & observability\n\nFine-tuning\n\nDeployment\n\nAI coding agents\n\nBut what I liked most...\n\nIt points you toward projects that require you to build systems.\n\nI was also happy to see that Louis-François Bouchard (@Whats_AI ) included our Agent Engineering: Building Multi-Agent Systems course as one of the recommended advanced resources.\n\nTo sum up:\n\nThe AI stack changes every few months.\n\nThe best investment isn't memorizing another framework...\n\nIt's building enough systems to understand the underlying patterns.\n\nIf you're starting your AI engineering journey or helping someone else get started, this is worth bookmarking.\n\nCheck it out here: https://github.com/louisfb01/start-ai-engineering\n\nP.S. Massive thank you to Louis-François Bouchard (@Whats_AI ) for putting this together and sharing my course."},"author":{"name":"Paul Iusztin","handle":"pauliusztin_","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/989690bdd97d74ea/avatar-4c78c916d3c6.jpg"},"metadata":{"likes":14,"shares":4,"views":980,"timestamp":"2026-07-02T12:30:00.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/yIjVEyhRne/media/0.mp4","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/yIjVEyhRne/media/thumb-0.jpg","type":"video","width":800,"height":406}],"aiComments":[]},{"id":"2072663035722777045","shareId":"ZwrzgQ1llH","platform":"x","url":"https://x.com/maarcoofdezz/status/2072663035722777045","shareMode":"full","createdAt":1783051760334,"postedAt":"2026-07-02T12:45:50.000Z","publishedDate":"2026-07-02T12:45:50.000Z","archivedAt":"2026-07-03T04:09:18.138Z","archivedDate":"2026-07-03T04:09:18.138Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/ZwrzgQ1llH","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2066540818421477376/img/qoTOXtlShDNFUVjT.jpg","contentType":"post","content":{"text":"Posiblemente este sea el mejor tutorial de Claude Cowork y en menos de 25 minutos.\n\nEs de Tina Huang (ex científica de datos en Meta).https://t.co/sZ9OXdV4VQ"},"author":{"name":"Marco","handle":"maarcoofdezz","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/0fb4144fe7172ca9/avatar-b6a831222033.jpg","bio":"21 | scaling brands with AI • founder @aiscwork | CMO @eComrads | powered by @rainbetcom"},"metadata":{"likes":281,"shares":28,"views":46433,"timestamp":"2026-07-02T12:45:50.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2066540818421477376/vid/avc1/1280x720/l8tqqLRjmcMlICDE.mp4?tag=28","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2066540818421477376/img/qoTOXtlShDNFUVjT.jpg","type":"video","width":1920,"height":1080,"duration":1371}],"aiComments":[]},{"id":"2072615317293907999","shareId":"8Ws2gFa258","platform":"x","url":"https://x.com/dunik_7/status/2072615317293907999","shareMode":"full","createdAt":1783051750677,"postedAt":"2026-07-02T09:36:13.000Z","publishedDate":"2026-07-02T09:36:13.000Z","archivedAt":"2026-07-03T04:09:08.106Z","archivedDate":"2026-07-03T04:09:08.106Z","viewCount":2,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/8Ws2gFa258","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072615102256173056/img/1K7qemMmzsBclvvQ.jpg","contentType":"post","content":{"text":"Someone just explained the entire agent stack - harness, loop engineering, evals, LLM ops, tracing - in 19 minutes.\n\nIt's clearer than most $500 courses, and it's free.\n\nharness → the structure around the model. loops → why the agent keeps going. evals → the \"no\" that catches it when it's wrong. tracing → how you find what actually broke.\n\nThat's the whole machine the article below breaks down, drawn out in one sitting.\n\nWatch it first to see the shape of it - then read the guide below to actually build one."},"author":{"name":"dunik","handle":"dunik_7","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/1e6f795f634d35c2/avatar-bcc8eeed3742.jpg","bio":"AI-powered wallet hunter | polymarket"},"metadata":{"likes":55,"shares":10,"views":3771,"timestamp":"2026-07-02T09:36:13.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2072615102256173056/vid/avc1/1114x720/9o9ru6Xlw5fhN9vH.mp4?tag=28","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072615102256173056/img/1K7qemMmzsBclvvQ.jpg","type":"video","width":1114,"height":720,"duration":1200}],"quotedPost":{"id":"2071584492804784468","platform":"x","url":"https://x.com/dunik_7/status/2071584492804784468","shareMode":"full","content":{"text":"Loop engineering: the skill that quietly ate prompt engineering\n\nhttp://x.com/i/article/2071531960934014976"},"author":{"name":"dunik","handle":"dunik_7","avatar":"https://pbs.twimg.com/profile_images/1874783436910571520/sHivdIrK_200x200.jpg"},"metadata":{"likes":62,"shares":11,"timestamp":"2026-06-29T13:20:05.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/F53WbVhUdu/quoted/0.jpg","type":"image","width":1200,"height":480},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/F53WbVhUdu/quoted/1.jpg","type":"image","width":640,"height":796},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/F53WbVhUdu/quoted/2.jpg","type":"image","width":1200,"height":750},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/F53WbVhUdu/quoted/3.jpg","type":"image","width":1200,"height":750}]},"aiComments":[]},{"id":"2072761875918455122","shareId":"ahQDukDu4J","platform":"x","url":"https://x.com/undefinedKi/status/2072761875918455122","shareMode":"full","createdAt":1783051727677,"postedAt":"2026-07-02T19:18:35.000Z","publishedDate":"2026-07-02T19:18:35.000Z","archivedAt":"2026-07-03T04:08:46.015Z","archivedDate":"2026-07-03T04:08:46.015Z","viewCount":2,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/ahQDukDu4J","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/C25cpmWT4M/media/thumb-0.jpg","contentType":"post","content":{"text":"Microsoft just released a full AI degree for free\n\nIt's called AI For Beginners, and it's a complete 12-week, 24-lesson curriculum. Free, on GitHub, no signup\n\nIt starts from zero and builds up. The history of AI, neural networks and deep learning, computer vision, natural language processing, transformers and large language models, even genetic algorithms and AI ethics.\n\nEvery lesson comes with reading, runnable Jupyter notebooks in both PyTorch and TensorFlow, quizzes, and hands-on labs. So you're not just reading theory, you're writing real code.\n\nThere's also a beginner track if you're totally new: \"Hello AI World,\" a simple neural network from scratch, an image classifier, and a sentiment analyzer.\n\nHere's the link: http://github.com/microsoft/AI-For-Beginners\n\nBookmark this"},"author":{"name":"Yarchi","handle":"undefinedKi","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/0cdd8407e08e8b24/avatar-44bcfc3142f9.jpg"},"metadata":{"likes":108,"shares":16,"views":14954,"timestamp":"2026-07-02T19:18:35.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/C25cpmWT4M/media/0.mp4","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/C25cpmWT4M/media/thumb-0.jpg","type":"video","width":720,"height":1280,"duration":27}],"quotedPost":{"id":"2070852381164630023","platform":"x","url":"https://x.com/undefinedKi/status/2070852381164630023","shareMode":"full","content":{"text":"10 Open-Source Repos That Quietly Make Claude Code 10x Better (Full Guide)\n\nhttp://x.com/i/article/2070302804812398592"},"author":{"name":"Yarchi","handle":"undefinedKi","avatar":"https://pbs.twimg.com/profile_images/2054664188417413120/DZKz0Tqn_200x200.jpg"},"metadata":{"likes":553,"shares":79,"timestamp":"2026-06-27T12:50:56.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/C25cpmWT4M/quoted/0.jpg","type":"image","width":1200,"height":480}]},"aiComments":[]},{"id":"2072767459908796782","shareId":"G4x6bsGtA8","platform":"x","url":"https://x.com/akshay_pachaar/status/2072767459908796782","shareMode":"full","createdAt":1783051720584,"postedAt":"2026-07-02T19:40:47.000Z","publishedDate":"2026-07-02T19:40:47.000Z","archivedAt":"2026-07-03T04:08:33.527Z","archivedDate":"2026-07-03T04:08:33.527Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/G4x6bsGtA8","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/PnfR5Xq9BB/media/thumb-0.jpg","contentType":"post","content":{"text":"RAG vs. Graph RAG vs. Agentic RAG, clearly explained!\n\nStandard RAG embeds documents into vectors and retrieves the most similar chunks via similarity search. For direct factual lookups, this works well.\n\nBut it breaks down when a query needs to connect facts spread across multiple documents. Similarity search retrieves individual chunks, not the relationships between them.\n\nGraph RAG adds a knowledge graph layer on top.\n\n→ During indexing, an LLM extracts entities and relationships from the documents.\n\n→ During retrieval, the system traverses these connections instead of relying on embedding similarity alone.\n\nThis is what enables multi-hop queries.\n\nSay a vector DB stores three facts about internal services:\n\n↳ \"The checkout service uses payments API.\"\n\n↳ \"The payments API runs on cluster-3.\"\n\n↳ \"Cluster-3 is scheduled for maintenance on Friday.\"\n\nSomeone asks: \"Will the checkout service be affected by Friday's maintenance?\"\n\nVector search can likely retrieve facts 1 and 3 because the query mentions \"checkout service\" and \"Friday maintenance.\"\n\nBut it will miss fact 2, which connects the payments API to cluster-3.\n\nThat middle fact sits too far from the query in embedding space. It mentions neither \"checkout\" nor \"maintenance,\" so it never makes it into the retrieved context.\n\nA knowledge graph connects these as linked entities, and graph traversal finds the full path in one query.\n\nAgentic RAG takes a different approach entirely.\n\nInstead of a fixed retrieval pipeline, an LLM agent decides at query time which tools to invoke, which sources to query, and in what order.\n\nCheck the visual below to understand the three architectures thoroughly.\n\nOne thing to note here is that these three aren't levels of sophistication that you need to graduate through.\n\nInstead, they solve different query types.\n\n↳ Single-hop factual lookups → standard RAG\n\n↳ Multi-hop relationship queries → Graph RAG\n\n↳ Dynamic multi-source tasks with tool use → Agentic RAG\n\n----\n\nEach of these architectures gets better when the underlying retrieval layer is efficient.\n\nI recently wrote about a new RAG approach that cuts corpus size by 40x, reduces tokens per query by 3x, and improves vector search relevance by 2.3x.\n\nThe article is quoted below."},"author":{"name":"Akshay 🚀","handle":"akshay_pachaar","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/a1223b06b0da3b6a/avatar-3debeed5ac88.jpg","bio":"Simplifying LLMs, AI Agents, RAG, and Machine Learning for you! • Co-founder @dailydoseofds_• BITS Pilani • 3 Patents • ex-AI Engineer @ LightningAI"},"metadata":{"likes":112,"shares":22,"views":13034,"timestamp":"2026-07-02T19:40:47.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/PnfR5Xq9BB/media/0.mp4","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/PnfR5Xq9BB/media/thumb-0.jpg","type":"video","width":2000,"height":1400}],"quotedPost":{"id":"2052743644411765230","platform":"x","url":"https://x.com/akshay_pachaar/status/2052743644411765230","shareMode":"full","content":{"text":"You're doing RAG wrong \n\nhttp://x.com/i/article/2052725748121956352"},"author":{"name":"Akshay 🚀","handle":"akshay_pachaar","avatar":"https://pbs.twimg.com/profile_images/1578327351544360960/YFpWSWIX_200x200.jpg"},"metadata":{"likes":598,"shares":65,"timestamp":"2026-05-08T13:33:17.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/PnfR5Xq9BB/quoted/0.jpg","type":"image","width":1200,"height":480},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/PnfR5Xq9BB/quoted/1.jpg","type":"image","width":1200,"height":670},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/PnfR5Xq9BB/quoted/2.jpg","type":"image","width":1200,"height":672},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/PnfR5Xq9BB/quoted/3.jpg","type":"image","width":1200,"height":670}]},"aiComments":[]},{"id":"2072687996151734680","shareId":"mHS9qZpuwY","platform":"x","url":"https://x.com/0xMovez/status/2072687996151734680","shareMode":"full","createdAt":1783051575004,"postedAt":"2026-07-02T14:25:01.000Z","publishedDate":"2026-07-02T14:25:01.000Z","archivedAt":"2026-07-03T04:06:11.398Z","archivedDate":"2026-07-03T04:06:11.398Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/mHS9qZpuwY","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072687569658056704/img/n3HbU1HrO09wXy5d.jpg","contentType":"post","content":{"text":"Head of Google Cloud AI:\n\n“the inner loop is capability - the agent investigates, implements, tests, reports.\n\nthe outer loop is agency - you decide, verify, approve, and own. That boundary is engineering.”\n\nin this 18-minute session, Director of Google Cloud AI gave the best breakdown on building agentic loops.\n\nWorth more than a $500 paid engineering course.\n\nWatch today, then explore how to build trading agent loops from scratch."},"author":{"name":"Movez","handle":"0xMovez","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/1c08021a94665113/avatar-5cf839236e7c.jpg","bio":"Content creator | AI researcher & agentic builder | @zscdao"},"metadata":{"likes":123,"shares":10,"views":11925,"timestamp":"2026-07-02T14:25:01.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2072687569658056704/vid/avc1/1280x720/nscsHtbyEnnU_rIi.mp4?tag=28","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072687569658056704/img/n3HbU1HrO09wXy5d.jpg","type":"video","width":1920,"height":1080,"duration":1079}],"quotedPost":{"id":"2072668933690126571","platform":"x","url":"https://x.com/0xMovez/status/2072668933690126571","shareMode":"full","content":{"text":"Loop Engineering for trading agents: 12 Steps to build a self-running Quant desk\n\nhttp://x.com/i/article/2071890135302426625"},"author":{"name":"Movez","handle":"0xMovez","avatar":"https://pbs.twimg.com/profile_images/1998148360478322688/851J4fBL_200x200.jpg"},"metadata":{"likes":213,"shares":37,"timestamp":"2026-07-02T13:09:16.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/BvwBbaVzqz/quoted/0.jpg","type":"image","width":1200,"height":480},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/BvwBbaVzqz/quoted/1.png","type":"image","width":688,"height":268},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/BvwBbaVzqz/quoted/2.png","type":"image","width":868,"height":774},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/BvwBbaVzqz/quoted/3.jpg","type":"image","width":1200,"height":667}]},"aiComments":[]},{"id":"2072681983826624878","shareId":"dwj7hacKeH","platform":"x","url":"https://x.com/Fluyeporlaweb/status/2072681983826624878","shareMode":"full","createdAt":1783045096107,"postedAt":"2026-07-02T14:01:08.000Z","publishedDate":"2026-07-02T14:01:08.000Z","archivedAt":"2026-07-03T02:18:13.154Z","archivedDate":"2026-07-03T02:18:13.154Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/dwj7hacKeH","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/5E1F0KckNl/media/thumb-0.jpg","contentType":"post","content":{"text":"alguien publicó en GitHub una CLI que le da a tu agente acceso a todo internet.\n\nsin API keys.\n\nsin cuentas.\n\nsin pagar nada.\n\n✅ lee tweets y timelines de Twitter\n\n✅ busca posts de Reddit y lee hilos completos con comentarios\n\n✅ extrae transcripciones de YouTube\n\n✅ navega repos de GitHub, abre issues y PRs\n\n✅ raspa XiaoHongShu y Bilibili\n\n✅ funciona con Claude Code, Cursor y Windsurf\n\npegas una línea.\n\ninstala solo.\n\ntu agente ya puede leer internet entero.\n\n47.7k estrellas. MIT. Gratis.\n\nse llama Agent Reach 👇"},"author":{"name":"PA13L0","handle":"Fluyeporlaweb","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/4aef53ce265b65e5/avatar-e69f4aef6d35.jpg","bio":"Dev y SEO • Creando activos digitales con IA"},"metadata":{"likes":364,"shares":56,"views":20666,"timestamp":"2026-07-02T14:01:08.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/5E1F0KckNl/media/0.mp4","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/5E1F0KckNl/media/thumb-0.jpg","type":"video","width":1892,"height":904,"duration":14}],"aiComments":[{"content":"**CLI (Command Line Interface, 명령줄 인터페이스)**\n터미널에서 텍스트 명령어를 입력해 프로그램을 실행하는 방식의 인터페이스. GUI(그래픽 화면) 없이 한 줄의 명령어만으로 설치·실행이 가능하다.\n\n**MCP (Model Context Protocol, 모델 컨텍스트 프로토콜)**\nAI 에이전트가 외부 도구(웹 스크래퍼, 데이터베이스, API 등)를 표준화된 방식으로 호출할 수 있도록 하는 오픈 JSON-RPC 2.0 기반 프로토콜. Claude Code, Cursor 등 여러 AI 클라이언트에서 지원한다. [출처](https://modelcontextprotocol.io/docs/getting-started/intro)\n\n**웹 스크래핑 (Web Scraping)**\n프로그램이 웹사이트에 자동으로 접근해 콘텐츠(텍스트, 이미지, 링크 등)를 수집·추출하는 기술. API 키 없이 플랫폼 데이터를 가져오는 Agent Reach의 핵심 동작 방식이다. [출처](https://brightdata.com/blog/ai/web-scraping-with-mcp)\n\n**Agent Reach**\nAI 에이전트에게 인터넷 전체 접근 권한을 부여하는 오픈소스 Python CLI 도구. Twitter, Reddit, YouTube, GitHub, 샤오홍수(小红书), Bilibili 등 13개 이상 플랫폼을 API 키 없이 읽고 검색할 수 있다. [출처](https://github.com/Panniantong/agent-reach)\n\n**MIT 라이선스 (MIT License)**\n소프트웨어를 자유롭게 사용·수정·배포할 수 있도록 허용하는 오픈소스 라이선스. 상업적 이용도 가능하며, 별도 비용 없이 누구나 활용할 수 있다.","meta":{"id":"ai-action-comment-aiaj_a11dee82-485e-4a04-9035-38dfbc5a6753","cli":"claude","model":"sonnet","type":"glossary","generatedAt":"2026-07-03T05:27:42.645Z"}}]},{"id":"2072693983008936088","shareId":"1J5QY5qEZ5","platform":"x","url":"https://x.com/mehmetsongur_/status/2072693983008936088","shareMode":"full","createdAt":1783045013623,"postedAt":"2026-07-02T14:48:48.000Z","publishedDate":"2026-07-02T14:48:48.000Z","archivedAt":"2026-07-03T02:16:51.291Z","archivedDate":"2026-07-03T02:16:51.291Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/1J5QY5qEZ5","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/RdZ7mXEGNJ/media/0.jpg","contentType":"post","content":{"text":"25 merkez bankası, 28 yıl, tek veri seti, üstelik açık kaynak\n\nGeorgia Tech’in World Central Bank (WCB) veri seti, 1996-2024 arası 25 merkez bankasının (TCMB dahil) politika metinlerini cümle düzeyinde topluyor. 380.200 cümle, 10 milyondan fazla kelime.\n\nBunların 25.000’i, 104 farklı etiketleyici tarafından üç görev için işaretlenmiş: duruş (şahin/güvercin/nötr), zaman perspektifi (ileriye dönük mü) ve belirsizlik.\n\n🔗 https://github.com/gtfintechlab/WorldCentralBanks"},"author":{"name":"Mehmet Songur","handle":"mehmetsongur_","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/19e23ccbcf945ee6/avatar-69fdf6507c4a.jpg","bio":"Doç. Dr./Assoc. Prof. Ph.D - Dicle Üniversitesi İktisat Bölümü | Dr./Ph.D. - Gazi Üniversitesi Data Analyst | AI"},"metadata":{"likes":119,"shares":14,"views":9735,"timestamp":"2026-07-02T14:48:48.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/RdZ7mXEGNJ/media/0.jpg","type":"image","width":1170,"height":612}],"aiComments":[]},{"id":"web-qkcqrg","shareId":"Zu5OBiCmv9","platform":"web","url":"https://www.langchain.com/blog/the-art-of-loop-engineering","title":"The Art of Loop Engineering","shareMode":"full","createdAt":1783045012193,"postedAt":"2026-06-16T17:00:00.000Z","publishedDate":"2026-06-16T17:00:00.000Z","archivedAt":"2026-07-03T02:16:50.083Z","archivedDate":"2026-07-03T02:16:50.083Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/Zu5OBiCmv9","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/z3vTaA4ea8/media/0.png","contentType":"article","content":{"text":"Agents are useful because they help us automate work by taking actions in the real world. But getting agents to do valuable work reliably takes more than just a good model: it requires a carefully designed harness that's fit to a set of tasks.\n\nThe core agent algorithm is simple: give the LLM context and let it call tools in a loop until it's done. This is the most fundamental loop. But it’s far from the only loop that powers agents. [Swyx](https://x.com/swyx) recently wrote a great piece on [\"loopcraft: the art of stacking loops\"](https://www.latent.space/p/ainews-loopcraft-the-art-of-stacking), the idea that you can stack and extend loops to build more effective agents.\n\nHere's how we think about that stack, and how to instrument each level with LangChain primitives.\n\n## Loop 1: The Agent\n\nAt its core, an agent is just a model calling tools in a loop until a task is complete.\n\n![Image](https://cdn.prod.website-files.com/65c81e88c254bb0f97633a71/6a317cf80046aaedb90c50c6_loop1%20(1).png)\n\nThis is what LangChain’s [`create_agent`](https://docs.langchain.com/oss/python/langchain/agents) gives you. Pick any model, plug in tools, and you have a working agent loop. [Tools](https://docs.langchain.com/oss/langchain/tools) are what give the agent the power to take action in the real world.\n\nTake our internal docs agent as an example (which we’ll use as a motivating example for the rest of this blog). At the first loop level, it receives a request for a documentation improvement, the model plans and draft changes, and it uses tools to clone repos, read files, write docs, open a pull request, etc.\n\n![Image](https://cdn.prod.website-files.com/65c81e88c254bb0f97633a71/6a317dca401c7eac8f267ab7_docs_writer_agent_loop_white_bg.png)\n\n## Level 2: Verification loop\n\nThe agent loop gets work done, but it doesn't always produce correct or consistent work on the first pass. When consistency matters, it's often useful to wrap it in a verification loop that checks the output and sends feedback back to the model when it falls short.\n\n![Image](https://cdn.prod.website-files.com/65c81e88c254bb0f97633a71/6a317d145d0b7b0f966909fc_loop2%20(1).png)\n\nThe verification loop adds a grader: something that checks the agent's output against a rubric and, if it fails, sends the result back with feedback. Graders can either be deterministic or agentic ([LLM as a judge](https://docs.langchain.com/langsmith/llm-as-judge) is a classic example, here).\n\n[`RubricMiddleware`](https://docs.langchain.com/oss/python/deepagents/rubric) handles this pattern, or you can wire it up with an [`after_agent` hook](https://docs.langchain.com/oss/langchain/middleware/overview) on `create_agent`.\n\nFor our docs writer example, the grader runs tests after each attempt, checking that all links resolve, all CI checks pass, and the diff is scoped to what was actually requested. No manual review needed to catch those classes of error.\n\nOne tradeoff: adding verification increases latency and cost per run. It's worth it when quality matters more than speed, which is most production use cases.\n\n![Image](https://cdn.prod.website-files.com/65c81e88c254bb0f97633a71/6a317dd76a131bd58de00fb9_docs_writer_verification_loop_white_bg.png)\n\n## Level 3: Event driven loop\n\nOne of the most important parts of agent development is the integrations layer: connecting your agent to your ecosystem so that it can run in the background.\n\nThe event-driven loop connects your agent to your ecosystem. An event fires — a new document lands, a schedule triggers, a webhook arrives — and the agent runs. The agent isn't something you invoke manually; it's a component running continuously inside a larger system.\n\n![Image](https://cdn.prod.website-files.com/65c81e88c254bb0f97633a71/6a317e69736b32a1e2363865_event_loop_generic_v3_white_bg.png)\n\n[LangSmith Deployment](https://info.langchain.com/agent-development-platform?utm_campaign=evergreen_agent_development_platform_cv&utm_campaign_id=23761370321&utm_ad_group_id=195261126163&utm_ad_id=805594028616&utm_network=g&utm_term=ai%20agent%20development%20platform&utm_campaign=evergreen_agents_cv&utm_source=google&utm_medium=cpc&hsa_acc=7906965105&hsa_cam=23761370321&hsa_grp=195261126163&hsa_ad=805594028616&hsa_src=g&hsa_tgt=kwd-2392721013549&hsa_kw=ai%20agent%20development%20platform&hsa_mt=p&hsa_net=adwords&hsa_ver=3&gad_source=1&gad_campaignid=23761370321&gbraid=0AAAAA-PkievTcIb-6awevyQxCB-9n-H6Z&gclid=CjwKCAjwxb7RBhA5EiwAQ-AAdF5XHKtTYLgQVrVYstdxYjTd0hcrCuqxvuACiKzOOdxcJdTza8HkwxoCDiQQAvD_BwE) supports the trigger infrastructure, including support for [cron schedules](https://docs.langchain.com/langsmith/cron-jobs) and [webhooks](https://docs.langchain.com/langsmith/use-webhooks). One popular example of crons in action is “heartbeats” in [openclaw](https://docs.openclaw.ai/gateway/heartbeat), which turn your agent into an always-on, proactive assistant.\n\nOur docs agent is powered by [Fleet](https://www.langchain.com/langsmith/fleet), our no-code agent builder. Fleet's [channels](https://docs.langchain.com/langsmith/fleet/channels) and [schedules](https://docs.langchain.com/langsmith/fleet/schedules) handle event-driven and cron-style triggers. We use a channel to fire off the docs agent whenever a message is sent in our `#docs-plz` Slack channel.\n\n![Image](https://cdn.prod.website-files.com/65c81e88c254bb0f97633a71/6a317e185240aaebb95cc7c3_docs_writer_event_loop_white_bg.png)\n\n## Level 4: Hill climbing loop\n\nThe first three loops automate work. The fourth (and arguably most important) automates improvement!\n\n![Image](https://cdn.prod.website-files.com/65c81e88c254bb0f97633a71/6a317e975f50e84a5bf17b80_hill_climbing_loop_generic_v2_white_bg.png)\n\nEvery agent run [produce a trace](https://docs.langchain.com/langsmith/observability): a record of what the model did, the tools it called, grader feedback, etc. Those traces contain high value signal regarding what's working and what isn't. The hill climbing loop runs an analysis agent over those traces and uses the findings to rewrite the harness with improved configuration. That can include prompt/tool tweaks or grader tweaks.\n\nIn LangSmith, you can use [Engine](https://www.langchain.com/langsmith/engine), our trace analysis agent, to instrument this fourth loop.\n\nWrapping up the docs agent analogy, we run engine over the docs agent traces to detect any issues. When multiple traces signal a potential problem, an issue is filed requesting changes to the offending prompt or tool.\n\n![Image](https://cdn.prod.website-files.com/65c81e88c254bb0f97633a71/6a317ea8af8c1790096c468d_docs_writer_hill_climbing_loop_v4_white_bg.png)\n\nThe key move here is that the return arrow doesn't just loop back to the top — it reaches inside and updates the agent loop directly. Each cycle of the outer loop makes the inner loops more effective.\n\n**Looking forward:** prompt and tool configuration are the most simple things to improve, but they're not the only options.For teams running open-weight models, the hill climbing loop can feed into RL fine-tuning, using trace or eval outcomes as training signal to improve the model itself.Auxiliary context like memory and retrieved skills can be improved the same way. The loop is the pattern; what it optimizes is up to you.\n\n## Human oversight and expertise\n\nAutomation doesn't mean removing humans from the loop. At every level, there are natural points where human oversight adds value. An automated grader can check whether links resolve; it takes a human to notice the framing is wrong for the audience. That kind of judgment, earned from context, experience, and taste, is exactly where human review earns its place.\n\nSome expertise should be codified in the prompt/tools themselves, but for sensitive actions, live human review is essential (think financial transactions, DB operations, etc). LangChain makes it straightforward to instrument these touch points in every loop:\n\n1. In the agent loop, [require human input](https://docs.langchain.com/oss/langchain/human-in-the-loop) before sensitive actions/tool calls\n2. In the verification loop, a human can act as the grader for sensitive workflows\n3. In the application loop, a human can approve outputs before they’re returned to the end user\n4. In the hill climbing loop, harness improvements can flow through human review before deployment All of LangChain’s open source frameworks make adding a “human in the loop” a [first class primitive](https://docs.langchain.com/oss/python/deepagents/human-in-the-loop).\n\n## Putting it all together\n\nIn case you’d prefer a more tabular view, here’s how those four loops stack together:\n\n| Loop | What it does | Impact | LangChain primitive |\n| --- | --- | --- | --- |\n| 1. Agent loop | Model calls tools repeatedly until a task is complete | Automate work | create_agent, any LangChain-supported model |\n| 2. Verification loop | Agent runs, output is scored against a rubric, retried with feedback if it fails | Ensure work quality and correctness | RubricMiddleware |\n| 3. Event driven loop | Events trigger agent runs that update a real system | Automated work at scale | LangSmith Deployment with cron triggers / webhooks or Fleet channels |\n| 4. Hill climbing loop | Traces from production runs feed an analysis agent that improves the harness config | Harness improvements | LangSmith Engine |\n\nThis is what loop engineering — or [loopcraft](https://www.latent.space/p/ainews-loopcraft-the-art-of-stacking), as swyx puts it — actually looks like in practice. AI leaders like [Steipete](https://x.com/steipete/status/2063697162748260627), [Boris](https://x.com/0xwhrrari/status/2064804504608887040), and [Andrej](https://www.youtube.com/watch?v=kwSVtQ7dziU) have all arrived at the same conclusion: the potential in agents is in the loops you build around them.\n\nWe’ve been thinking about loops 1 and 2 for a while. But focus should pivot to loops 3 and 4 where value compounds by embedding agents into your ecosystem that continuously improve in response to your criteria.\n\nSatya [frames the organizational stakes](https://x.com/satyanadella/status/2066182223213293753): companies that build learning loops early,where human judgment and token capital compound together, will build an advantage that's hard to replicate.\n\n## Acknowledgements\n\nThanks to [Vivek](https://x.com/Vtrivedy10), [Mason](https://x.com/masondrxy), [Harrison](https://x.com/hwchase17?lang=en), and [Hunter](https://x.com/huntlovell) for thoughtful review.\n\n## Reference\n\n- [deepagents quickstart](https://docs.langchain.com/oss/python/deepagents/quickstart)\n- [create_agent docs](https://docs.langchain.com/oss/python/langchain/agents)\n- [rubric middleware](https://docs.langchain.com/oss/python/deepagents/rubric)\n- [cron jobs](https://docs.langchain.com/langsmith/cron-jobs), [webhooks](https://docs.langchain.com/langsmith/use-webhooks)\n- [langsmith engine](https://www.langchain.com/langsmith/engine)\n- [fleet channels](https://docs.langchain.com/langsmith/fleet/channels)"},"author":{"name":"LangChain Accounts","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/de3346441671c839/avatar-5b84dffca269.png","bio":"Agents automate real-world work, but reliable performance requires more than a good model, it requires a carefully designed harness built for specific tasks. This post explores the core agent loop, how stacking and extending loops builds more effective agents, and how to instrument each level with LangChain primitives."},"metadata":{"timestamp":"2026-06-16T17:00:00.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/z3vTaA4ea8/media/0.png","type":"image","width":1980,"height":1114}],"aiComments":[{"content":"**에이전트 하네스 (Agent Harness)**\nLLM을 실제로 작동하는 에이전트로 만들어주는 소프트웨어 인프라. 루프, 도구 호출, 컨텍스트 관리, 메모리, 가드레일, 트레이싱 등 에이전트를 감싸는 모든 scaffolding을 포함한다. [출처](https://www.databricks.com/blog/ai-harness)\n\n**루브릭 (Rubric)**\n에이전트 출력물의 품질을 평가하기 위한 명시적 기준 체계. 단일 점수 대신 여러 차원(링크 유효성, CI 통과 여부, 범위 준수 등)으로 세분화된 채점 기준을 정의하며, 검증 루프의 채점기(grader)가 이를 기반으로 재시도 여부를 판단한다. [출처](https://www.confident-ai.com/blog/definitive-ai-agent-evaluation-guide)\n\n**LLM as a Judge**\n하나의 LLM이 다른 LLM의 출력을 평가하는 방식. 결정론적 규칙으로 판단하기 어려운 품질(문체, 논리적 일관성, 맥락 적합성 등)을 자동으로 채점할 때 사용하는 에이전틱 채점 패턴이다. [출처](https://huggingface.co/blog/agent-glossary)\n\n**트레이스 (Trace)**\n에이전트 실행 중 모델이 수행한 모든 행동의 기록 — 호출한 도구, 중간 추론 과정, 채점기 피드백 등을 포함한다. 힐 클라이밍 루프의 분석 에이전트가 이 트레이스를 입력으로 받아 하네스 개선점을 도출한다. [출처](https://blog.jetbrains.com/pycharm/2026/05/llm-evaluation-and-ai-observability-for-agent-monitoring/)\n\n**힐 클라이밍 루프 (Hill Climbing Loop)**\n프로덕션 실행에서 수집된 트레이스를 분석 에이전트에게 공급하여 프롬프트·도구 설정 등 하네스 구성을 자동으로 개선하는 메타 루프. 각 사이클마다 내부 루프의 성능이 향상되는 자기 강화 구조를 가진다. [출처](https://www.langchain.com/blog/the-art-of-loop-engineering)","meta":{"id":"ai-action-comment-aiaj_5726bc80-9d3b-4e26-ba13-859f8ca23804","cli":"claude","model":"sonnet","type":"glossary","generatedAt":"2026-07-03T03:33:46.404Z"}}]},{"id":"2072730369044988137","shareId":"dpf582MlVt","platform":"x","url":"https://x.com/igus_ai/status/2072730369044988137","shareMode":"full","createdAt":1783044937452,"postedAt":"2026-07-02T17:13:24.000Z","publishedDate":"2026-07-02T17:13:24.000Z","archivedAt":"2026-07-03T02:15:32.423Z","archivedDate":"2026-07-03T02:15:32.423Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/dpf582MlVt","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2065399693090095104/img/jg9D6mnEBBQgXGbn.jpg","contentType":"post","content":{"text":"ESTA TRADER CHINA USÓ CLAUDE FABLE 5 PARA CONSTRUIR UN BOT DE TRADING\n\nEn este tutorial enseña cómo replicarlo desde 0 paso por paso\n\nEn tan solo 31 minutos. 100% gratis.\n\nLo he subtitulado en español e inglés.\n\nGuárdalo antes de que lo bloqueen 🔖 https://t.co/FNmVxBCttI"},"author":{"name":"gus","handle":"igus_ai","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/0bc9d8aa9ce9ff2f/avatar-c369747e20f1.jpg","bio":"ex @microsoft, now i talk about tech, trading and AI. growth @aiscwork"},"metadata":{"likes":343,"shares":75,"views":39577,"timestamp":"2026-07-02T17:13:24.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2065399693090095104/vid/avc1/1280x720/MB2WBdEHPy_cbgdd.mp4?tag=27","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2065399693090095104/img/jg9D6mnEBBQgXGbn.jpg","type":"video","width":2560,"height":1440,"duration":1879}],"aiComments":[]},{"id":"2072706105885552959","shareId":"XH999OqQ7D","platform":"x","url":"https://x.com/LukeberryPi/status/2072706105885552959","shareMode":"full","createdAt":1783044738862,"postedAt":"2026-07-02T15:36:59.000Z","publishedDate":"2026-07-02T15:36:59.000Z","archivedAt":"2026-07-03T02:12:13.627Z","archivedDate":"2026-07-03T02:12:13.627Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/XH999OqQ7D","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/AIZ8BMU5DY/media/0.jpg","contentType":"post","content":{"text":"essa skill /to-issues do Matt Pocock é realmente MUITO BOA\n\nsabe quando você refinou, respondeu perguntar do /grill-me e agora tá com um plano gigante na sua frente?\n\nse você jogar o agente nesse plano, a chance de sair caquinha é grande, por um motivo: você vai entupir o contexto tentando fazer tudo de uma vez\n\nai que entra /to-issues, ele quebra o plano em pequenas issues com a tag \"ready for agent\"\n\naí o agente consegue iterar em cada uma, sabe exatamente qual é o AC e consegue verificar por si só (ou spawnar um subagent para revisar)\n\nessa história só fica melhor :)"},"author":{"name":"Luke Berry","handle":"LukeberryPi","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/9fa060d07b0bf2a2/avatar-ae14fb35585e.jpg","bio":"your favourite developer's favorite dj @raycast ambassador 🇧🇷🇬🇧"},"metadata":{"likes":468,"shares":15,"views":18805,"timestamp":"2026-07-02T15:36:59.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/AIZ8BMU5DY/media/0.jpg","type":"image","width":1412,"height":1162},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/AIZ8BMU5DY/media/1.jpg","type":"image","width":2598,"height":1944}],"aiComments":[{"content":"**/to-issues**\nMatt Pocock이 만든 Claude Code 스킬로, 완성된 개발 계획(PRD)을 즉시 실행 가능한 작은 GitHub 이슈들로 분해해주는 도구. 각 이슈에는 범위, 의존성, 인수 조건이 포함되어 에이전트가 독립적으로 처리할 수 있도록 한다. [출처](https://github.com/mattpocock/skills)\n\n**/grill-me**\n구현 시작 전 설계 결정을 구조적으로 검토하는 Claude Code 스킬. 숨겨진 가정, 엣지 케이스, 미해결 요구사항을 질문을 통해 표면화하여 에이전트에게 명확한 계획을 넘기기 전 정제 과정을 거치게 한다. [출처](https://crossaitools.com/skills/mattpocock/skills/grill-me)\n\n**AC (Acceptance Criteria, 인수 조건)**\n소프트웨어 개발에서 특정 기능이나 이슈가 \"완료\"로 간주되기 위해 충족해야 하는 조건의 집합. 에이전트가 작업 완료 여부를 스스로 검증하는 기준으로 활용된다.\n\n**컨텍스트 (Context Window)**\nLLM(대형 언어 모델)이 한 번의 요청에서 처리할 수 있는 텍스트의 최대 용량. 하나의 에이전트 세션에 너무 많은 작업을 몰아넣으면 컨텍스트가 포화되어 결과물 품질이 저하된다. [출처](https://medium.com/@anmjawad007/the-two-matt-pocock-skills-i-use-in-almost-every-ai-coding-session-57c468938145)\n\n**서브에이전트 (Subagent)**\n메인 에이전트가 특정 하위 작업(예: 코드 리뷰)을 처리하기 위해 별도로 생성하는 독립적인 AI 에이전트. 작업을 분리함으로써 각 에이전트의 컨텍스트를 깨끗하게 유지할 수 있다.","meta":{"id":"ai-action-comment-aiaj_370d2b9e-f833-415b-af24-1a41583c09fe","cli":"claude","type":"glossary","generatedAt":"2026-07-03T02:13:58.262Z"}}]},{"id":"2072349006710337972","shareId":"cyU7IR1bh6","platform":"x","url":"https://x.com/XAMTO_AI/status/2072349006710337972","shareMode":"full","createdAt":1783044555047,"postedAt":"2026-07-01T15:58:00.000Z","publishedDate":"2026-07-01T15:58:00.000Z","archivedAt":"2026-07-03T02:09:10.504Z","archivedDate":"2026-07-03T02:09:10.504Z","viewCount":3,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/cyU7IR1bh6","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/ZE6vQbKfeG/media/0.png","contentType":"post","content":{"text":"GitHub上有个 39.7k Star 的项目叫 Sniffnet，它厉害的地方在于把网络监控这件事，做到了普通人也能轻松看懂的程度。\n\n很多人第一次打开Wireshark时的反应，基本都是：满屏的专业术语和数据包，直接劝退。但 Sniffnet 的体验完全不同，它能实时显示上传下载流量、当前的连接主机、访问的域名、IP 地理位置，甚至连哪个软件在偷偷联网都能让你看个明白。它还支持流量过滤、PCAP文件导入导出，以及多网卡监控这些专业功能。\n\n界面设计得非常直观友好，用起来很舒服。你完全不用去研究复杂的抓包规则，也无需记忆繁琐的命令行参数，打开软件就能一眼看明白你的电脑究竟在和哪些外部设备通信。很多用户直接把它当成 GlassWire 的免费替代品来用。\n\nSniffnet 支持 Windows、macOS、Linux 全平台，使用 Rust 语言编写，资源占用极低。对于经常折腾代理、NAS、Docker 和服务器的朋友来说，这绝对是一款称手的装机必备工具。\n\n🔗 https://github.com/GyulyVGC/sniffnet"},"author":{"name":"Amto","handle":"XAMTO_AI","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/d8546ce0e0dc0123/avatar-9644dceec29c.jpg"},"metadata":{"likes":367,"shares":84,"views":24130,"timestamp":"2026-07-01T15:58:00.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ZE6vQbKfeG/media/0.png","type":"image","width":864,"height":782}],"aiComments":[{"content":"**PCAP (패킷 캡처 파일)**\n네트워크상에서 캡처된 패킷 데이터를 저장하는 파일 형식으로, OSI 레이어 2~7의 모든 바이트를 그대로 담고 있습니다. tcpdump 및 libpcap 라이브러리와 함께 1990년대에 개발되었으며, 네트워크 트래픽 분석의 표준 포맷으로 사용됩니다. [출처](https://www.endace.com/learn/what-is-a-pcap-file)\n\n**Wireshark**\n오픈소스 네트워크 프로토콜 분석기로, 실시간 네트워크 트래픽을 캡처하고 패킷 단위로 상세하게 표시해주는 도구입니다. 세계에서 가장 널리 쓰이는 패킷 분석 툴이지만 전문적인 지식이 없으면 활용하기 어렵습니다. [출처](https://www.comparitech.com/net-admin/pcap-guide/)\n\n**GlassWire**\n5,300만 회 이상 다운로드된 상용 개인 방화벽 겸 네트워크 모니터링 도구입니다. 네트워크 활동을 시각적인 그래프로 표시하고 이상 트래픽 발생 시 알림을 제공하며, Sniffnet은 이 도구의 무료 대체재로 자주 언급됩니다. [출처](https://www.glasswire.com/features/)\n\n**NAS (Network Attached Storage, 네트워크 결합 스토리지)**\n네트워크에 직접 연결되어 여러 기기에서 파일을 공유·접근할 수 있도록 하는 전용 스토리지 장치입니다. 가정이나 소규모 기업에서 서버처럼 활용되며, 자체적인 운영체제와 네트워크 인터페이스를 갖추고 있습니다.\n\n**Rust (러스트)**\n성능, 메모리 안전성, 동시성을 핵심 목표로 설계된 시스템 프로그래밍 언어입니다. 가비지 컬렉터 없이도 메모리 안전을 보장하는 독자적인 소유권 시스템을 갖추고 있으며, 2016년부터 2025년까지 10년 연속 \"가장 선호하는 프로그래밍 언어\"로 선정되었습니다. Sniffnet이 Rust로 작성된 덕분에 리소스 점유율이 매우 낮습니다. [출처](https://en.wikipedia.org/wiki/Rust_(programming_language))","meta":{"id":"ai-action-comment-aiaj_3035190d-3bca-40a5-a469-b9ac254254c7","cli":"claude","model":"sonnet","type":"glossary","generatedAt":"2026-07-03T03:37:30.520Z"}}]},{"id":"2072605320132575489","shareId":"WIvBsKMSy0","platform":"x","url":"https://x.com/WaldronLewis/status/2072605320132575489","title":"일론 머스크는 Y Combinator에서 45분 동안 직접 만나 세상 어떤 경영대학원보다도 더 효과적으로 세상을 바꿀 기업을 만드는 방법을 설명했습니다. 그...","shareMode":"full","createdAt":1783043863381,"postedAt":"2026-07-02T08:56:30.000Z","publishedDate":"2026-07-02T08:56:30.000Z","archivedAt":"2026-07-03T01:57:39.695Z","archivedDate":"2026-07-03T01:57:39.695Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/WIvBsKMSy0","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/HOUJu5WpOd/media/0.jpg","contentType":"article","content":{"text":"# 일론 머스크는 Y Combinator에서 45분 동안 직접 만나 세상 어떤 경영대학원보다도 더 효과적으로 세상을 바꿀 기업을 만드는 방법을 설명했습니다. 그...\n\n## 1\n\n일론 머스크는 Y Combinator에서 45분 동안 직접 만나 세상 어떤 경영대학원보다도 더 효과적으로 세상을 바꿀 기업을 만드는 방법을 설명했습니다. 그가 젊은 창업가들로 가득 찬 방에서 전한 조언은 다음과 같습니다.\n\n1. 위대한 것을 만들려고 하지 마세요. 유용한 것을 만들려고 노력하세요.\n\n모두가 위대함에 집착합니다. 하지만 머스크는 그것이 잘못된 목표라고 말합니다. \"저는 애초에 위대한 것을 만들겠다고 생각하지 않았습니다. 유용한 것을 만들어보고 싶었습니다. 특별히 훌륭한 것을 만들 거라고는 생각하지 않았죠. 가능성은 낮아 보였지만, 적어도 시도는 해보고 싶었습니다.\" 먼저 유용함을 목표로 삼으세요. 위대함은 따라오는 부산물일 뿐입니다.\n\n2. 정문으로 들어갈 수 없다면, 직접 문을 만들어라.\n\n머스크는 첫 회사를 창업하기 전 넷스케이프에 취직하려고 시도했었다. \"넷스케이프에 이력서를 보냈는데 아무도 연락을 주지 않았어요. 로비에서 누군가를 만날까 기대하며 서성거려 봤지만, 너무 수줍어서 말을 걸 용기가 나지 않았죠. 그래서 '이건 말도 안 돼, 내가 직접 소프트웨어를 만들어야겠다'라고 생각했어요.\" 그는 창업자가 되려고 했던 것은 아니었다. 아무도 그를 고용하지 않았기 때문에 창업자가 된 것이다.\n\n3. 그는 사무실에서 잠을 자고 YMCA에서 샤워를 했다.\n\n그의 첫 회사의 시작은 화려하지 않았다. \"숙소를 구할 돈조차 없었어요. 사무실 월세가 500달러였기 때문에 사무실에서 자고 YMCA에서 샤워를 했죠.\" 제대로 된 인터넷도 사용할 형편이 안 돼서 사무실 바닥에 구멍을 뚫고 아래층 인터넷 업체까지 케이블을 연결했다고 한다. 바로 그 사람이 훗날 세계에서 가장 부유한 사람이 될 인물의 창업자였다.\n\n4. 승부를 걸지 마라.\n\n머스크는 첫 회사를 매각했을 때 2천만 달러짜리 수표를 받았습니다. 그의 은행 잔고는 하룻밤 사이에 1만 달러에서 2천만 달러로 급증했습니다. 대부분의 사람이라면 거기서 멈췄을 것입니다. 하지만 그는 거의 모든 돈을 다음 회사에 재투자했습니다. \"저는 기회를 놓치지 않았습니다.\" 그는 수십 년 후에도 같은 방식으로 여러 번 반복했습니다. 그는 돈이 놀고 있는 것을 싫어합니다. 돈은 다음 목표를 향한 원동력입니다.\n\n5. 먼저 사명을 정하고, 그 사명을 사업화하는 방법을 역으로 생각해 보세요.\n\n머스크는 돈을 벌기 위해 스페이스X를 설립한 것이 아닙니다. 그는 NASA 웹사이트에 들어가 인간이 언제 화성에 갈 계획인지 알아봤지만, 아무런 계획도 없었습니다. 그래서 그는 직접 계획을 세우기로 했습니다. \"이전에 로켓 스타트업이 성공한 사례는 없었습니다. 작은 가능성이라도 없는 것보다는 낫죠.\" 임무가 우선이었고, 사업 모델은 그 후에 만들어졌습니다.\n\n6. 그는 실패할 것을 예상하고 SpaceX를 시작했습니다.\n\n그는 실패 확률에 대해 매우 솔직하게 이야기합니다. \"SpaceX는 2002년 중반에 실패할 것을 예상하고 시작했습니다. 아마 90% 확률로 실패했을 겁니다. 사람들을 모집할 때도 '우리는 망할 가능성이 높지만, 아주 작은 가능성으로 살아남을 수도 있다'고 말했죠.\" 처음 세 번의 발사는 실패했습니다. 네 번째 발사는 자금이 바닥난 상태에서 성공했습니다. \"네 번째 발사마저 실패했다면 완전히 끝장이었을 겁니다. 정말 아슬아슬하게 살아남았죠.\"\n\n7. 모든 문제를 물리학 원리에 기반하여 분석하라.\n\n이것이 바로 머스크의 사고방식의 핵심입니다. \"제1원칙이란 사물을 가장 기본적인 요소로 분해한 다음, 거기서부터 추론해 나가는 것을 의미합니다. 유추에 의한 추론과는 정반대죠.\" 그의 예시는 로켓입니다. 모두가 과거 로켓의 가격을 기준으로 새 로켓의 가격을 책정했습니다. 하지만 머스크는 로켓이 실제로 무엇으로 만들어졌는지, 원자재 가격을 알아본 결과 재료비가 과거 가격의 1~2%에 불과하다는 사실을 발견했습니다. 나머지는 비효율적인 부분이었고, 그는 이를 개선할 수 있었습니다.\n\n8. 어떤 일이 24개월 걸린다는 말을 들으면, 그 일을 6개월 단위로 쪼개서 하세요.\n\n작년에 xAI는 자사 AI를 훈련시키기 위해 거대한 컴퓨터가 필요했습니다. 공급업체들은 18개월에서 24개월이 걸릴 것이라고 했습니다. \"하지만 우리는 6개월 안에 그걸 해내야 경쟁력을 유지할 수 있을 겁니다.\" 그래서 그는 컴퓨터를 여러 부분으로 나누었습니다. 건물이 필요해서 낡은 공장을 찾았고, 전력이 필요해서 발전기를 빌렸습니다. 냉각이 필요해서 미국 내 이동식 냉각 설비 용량의 4분의 1에 달하는 장비를 임대했습니다. 그는 데이터 센터에서 잠을 자고 직접 케이블을 설치했습니다. 결국 모든 것이 완성되었습니다.\n\n9. 자존심과 능력의 비율을 잘 살펴보세요.\n\n머스크가 젊은 창업가들에게 주는 가장 날카로운 조언은 바로 자기 자신에게 솔직하라는 것입니다. \"가장 흔한 실패 원인은 자만심이 능력보다 너무 커지는 것입니다. 그러면 현실과의 피드백 고리가 끊어지게 되죠.\" 자만심은 작게 유지하고, 모든 것에 대한 책임을 내면화하며, 진실에 냉철하게 집중해야 합니다. \"현실과의 연결 고리를 확실하게 닫아야 합니다. 그게 정말 중요한 부분입니다.\"\n\n10. 명예가 아닌 일을 쫓으십시오.\n\n그의 마지막 철학은 모든 것을 하나로 엮어줍니다. \"유용한 사람이 되는 것은 정말 어렵습니다. 총 효용 곡선 아래 면적은 당신이 다른 사람들에게 얼마나 유용하게 기여했는지에 대한 곱입니다. 진정한 일을 하기를 열망한다면 성공 확률이 훨씬 높아집니다. 영광을 열망하지 말고, 일을 열망하십시오.\"\n\n그는 수년간 조롱을 받았습니다. 언론은 그를 \"인터넷에서 로켓 회사를 차리려는 녀석\"이라고 불렀습니다. 그 역시 그 말이 터무니없다는 데 동의했습니다. 하지만 그는 작은 가능성이라도 전혀 없는 것보다는 낫다고 생각하여 결국 해냈습니다.\n\n그런데 문제는...\n\n머스크는 모든 행보가 공개되기 때문에 현존하는 창업가 중 가장 많은 팔로워를 보유하게 되었습니다. 제품 출시, 실패, 그리고 이번 강연과 같은 모든 것들이 공개됩니다. 그가 소유한 회사들은 그를 강력하게 만들었고, 그의 개인 브랜드는 그가 말을 마치기도 전에 그의 모든 말이 전 세계로 퍼져나가게 했습니다.\n\n저희는 고객이 손 하나 까딱하지 않고도 X 플랫폼 및 그 외 플랫폼에서 대규모 유통망을 구축하고 개인 브랜드를 성장시켜 드립니다.\n\n창업자 또는 벤처캐피탈리스트로서 이러한 홍보 기회를 찾고 계신다면 아래에서 상담 예약을 하세요.\n\n저희 콘텐츠는 주 평균 150만 조회수를 기록하고 있습니다.\n\nhttps://translate.google.com/\n\n{{IMAGE_0}}\n\n---\n\n## 2\n\nCredit:"},"author":{"name":"Lewis","handle":"WaldronLewis","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/8723b6549e23668c/avatar-9f6fb861f210.jpg","bio":"Co-Founder of UDG - We build viral content infrastructure for brands - 1B views to date"},"metadata":{"timestamp":"2026-07-02T08:56:30.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/HOUJu5WpOd/media/0.jpg","type":"image","width":640,"height":360}],"comments":[{"id":"2072846707126116484","content":"Excellent startup advice, you're still hoarding way too much money though it should be distributed among the lower classes to help provide them more opportunities to climb the social ladder, IMO","author":{"name":"SOVEREIGN","username":"munreader","handle":"munreader","url":"https://x.com/munreader"},"timestamp":"2026-07-03T00:55:41.000Z"},{"id":"2072819696618135770","content":"https://\nx.com/dovlan1990/sta\ntus/2072814053643391230?s=20\n…\n\nТолько я хз как теперь относиться","author":{"name":"Nikita","username":"dovlan1990","handle":"dovlan1990","url":"https://x.com/dovlan1990"},"timestamp":"2026-07-02T23:08:21.000Z"},{"id":"2072794629305684081","content":"Cool I will look over this later","author":{"name":"Emma Leonhart","username":"EmmaNoMikoto","handle":"EmmaNoMikoto","url":"https://x.com/EmmaNoMikoto"},"timestamp":"2026-07-02T21:28:44.000Z"},{"id":"2072835888132841670","content":"Yeah why would somebody would sit down with me for 45 minutes","author":{"name":"Danny Lynn","username":"DannyLy82999744","handle":"DannyLy82999744","url":"https://x.com/DannyLy82999744"},"timestamp":"2026-07-03T00:12:41.000Z"},{"id":"2072832855533609382","content":"@readwise\n save thread","author":{"name":"Agust Olafsson","username":"agustolafsson","handle":"agustolafsson","url":"https://x.com/agustolafsson"},"timestamp":"2026-07-03T00:00:38.000Z"},{"id":"2072754253589017028","content":"@lewis\n completely agree with 1","author":{"name":"Asrar","username":"asrrrrrr217","handle":"asrrrrrr217","url":"https://x.com/asrrrrrr217"},"timestamp":"2026-07-02T18:48:18.000Z"},{"id":"2072855046287724884","content":"Don't aspire to glory, aspire to work - that's his best statement. I had an Illuminati woman try to recruit me, pitching I would have glory. I was astonished how dumb she was to think I'd go for that. God bless Elon more.","author":{"name":"Skip Press","username":"skippress","handle":"skippress","url":"https://x.com/skippress"},"timestamp":"2026-07-03T01:28:49.000Z"}],"aiComments":[]},{"id":"2072842101067481210","shareId":"i1S6nuTwr0","platform":"x","url":"https://x.com/shao__meng/status/2072842101067481210","shareMode":"full","createdAt":1783043855728,"postedAt":"2026-07-03T00:37:23.000Z","publishedDate":"2026-07-03T00:37:23.000Z","archivedAt":"2026-07-03T01:57:30.763Z","archivedDate":"2026-07-03T01:57:30.763Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/i1S6nuTwr0","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/FgQdlzbQ88/media/0.jpg","contentType":"article","content":{"text":"Every 团队使用 Codex 的深度实践\n\nhttps://every.to/context-window/codex-in-practice?utm_source=X\n\n# 背景不同的五人、五种不同的工作流\n\n① Natalia：非技术构建者的“低摩擦 Claude Code”\n\n· 痛点：她曾在 Claude Code 中精心维护文件夹结构，但在 Codex 里无需自己搭建。\n\n· 用法：每天打开当天优先的项目线程，让 Codex 自行决定架构与文件组织。\n\n· 关键场景：用 CRM（Attio）管理客户关系时，她给 Codex 访问邮箱、会议记录和销售管线逻辑，让它在夜间自动 enrich 数百条客户记录——原本需要数周的手工工作。\n\n· 个人应用：为父亲的多护士护理流程建立“家庭操作系统”，把分散的医疗预约、随访协议、家属信息整合到一个中心位置。\n\n启示：Codex 对非技术用户的核心价值是降低“系统搭建”的认知负担，把“架构能力”外包给模型。\n\n② Dan：长线程 + 内置浏览器 + 路由线程\n\n· 原则：让 Codex 获得完成某任务所需的全部上下文。\n\n· 长线程（long-running threads）：为邮件处理、Slack/会议摘要、招聘等重复任务建立不重置的线程，保留完整历史。\n\n· 内置浏览器：Codex 可在应用内浏览网页、点击交互，无需用户复制粘贴上下文。这对调试 UI 和知识工作都有效。\n\n· 路由线程（router thread）：他做了一个小工具 Mailroom，给 Codex 一个专属邮箱；路由线程每几分钟检查收件箱，把请求分发到对应的专业线程（合同→法律线程，权限申请→内部运营线程）。\n\n启示：这是目前最成熟的“多智能体”个人工作流雏形——通过线程分工实现规模化任务处理。\n\n③ Katie：本地“上下文文件夹”系统\n\n· 背景：Claude Code 重度用户，迁移到 Codex 后让模型采访自己，自动构建更适合 agent 的文件系统。\n\n· Katie Context 文件夹包含：\n\n · AGENTS.md：目录与使用说明\n\n · identity：角色与项目类型\n\n · preferences：输出格式偏好，例如需要可分享时创建 Proof 文档\n\n · rules：“不要这样做”清单\n\n · project map：各工作领域对应的文件、指令、流程\n\n · Voice.md：写作风格、语言、语气指南\n\n· 自动化创意库：她为专栏 Working Overtime 建立“idea farm”，让 Codex 监控 Slack 频道，按她的风格与选题标准评分并添加编辑备注。\n\n启示：对于需要稳定输出质量的知识工作者，显式化个人偏好与风格比单纯依赖模型记忆更可靠。\n\n④ Austin：结果导向的轻量模式\n\n· 风格：自认不擅长组织，因此避免复杂结构。\n\n· 用法：只给 Codex 一个结果目标，连接相关信息源，让它直接开工。\n\n· 例子：为活动制作后续内容包时，他丢入 transcript、聊天记录、录音、Monologue 语音备忘录，让 Codex 去 Notion 和 Slack 搜索、建立 GitHub 资源库、起草邮件。\n\n· 迭代机制：如果初稿不好，他让 Codex 基于既有机构上下文自我审计，找出哪里出错并改进；遇到增长工作问题时，让 Codex 审计 Every GrowthOS 仓库，删除过时指令。\n\n启示：反馈-审计-精简的循环，比一次完美提示更重要。适合不愿花精力维护系统的人。\n\n⑤ Kieran：可移植的跨 Agent 上下文系统\n\n· 定位：Codex 只是他个人 AI 系统中的一个工具，核心资产是一个跨设备同步的上下文文件夹。\n\n· 输入源：会议记录、Monologue 语音笔记、日记、Limitless 吊坠录音。\n\n· 自动处理：新素材进入后，自动工作流会分类、提取关键信息并写入对应文件。\n\n· 记忆层：生成日/周/月摘要，成为所有 agent 可读的可复用记忆。\n\n· Codex 的角色：当需要跨工具快速搜索和执行时调用，例如收集 Cora 早期申请者的联系方式、补全缺失邮箱、建表、邀请进 Slack、发欢迎邮件。\n\n启示：上下文资产的可移植性比绑定单一 agent 更重要。Codex 是执行层，而非记忆中枢。\n\n# 跨案例的共性模式\n\n从五个人的实践中可以提炼出四条普适原则：\n\n1. 上下文比提示更重要\n\n无论是长线程、文件夹系统，还是可移植知识库，成功的关键都在于把 Codex 接入邮件、会议、Slack、CRM、浏览器等真实信息源。\n\n2. 让 Codex 自己设计系统\n\n多数人并非手动搭建文件夹结构，而是让 Codex 通过“采访我”来生成适合自己的架构。这降低了启动门槛。\n\n3. 把重复工作委托为“后台任务”\n\nNatalia 的 CRM enrich、Kieran 的申请者处理、Dan 的邮件路由，共同点是：把低判断、高耗时的任务变成异步执行。\n\n4. 建立审计与反馈循环\n\nAustin 的“审计哪里出错”和 Katie 的“规则文件”都说明，Codex 不是一次设定就结束，需要持续根据错误更新约束。\n\n# 实用建议：如何开始\n\n文章本身的结论最值得参考：不要过度思考，先选一个风格切入。\n\n· 如果你不想搭建系统 → 用 Austin 的结果导向法\n\n· 如果你写作为主、重视风格一致性 → 用 Katie 的上下文文件夹法\n\n· 如果你处理大量重复性行政/运营任务 → 用 Kieran 的可移植上下文法\n\n· 如果你想让 Codex 像员工一样管理多线程工作 → 用 Dan 的长线程 + 路由法\n\n· 如果你是非技术背景、希望降低工具学习成本 → 用 Natalia 的低摩擦 Claude Code 法"},"author":{"name":"meng shao","handle":"shao__meng","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/1e9f6320361186f5/avatar-25bb4f91cb4f.jpg","bio":"Building AI Agents for design & media. 分享新产品、开源项目，以及 AI 创业与职场观察。公众号 / 小红书：AI 启蒙小伙伴｜合作请私信"},"metadata":{"likes":4,"shares":0,"views":1061,"timestamp":"2026-07-03T00:37:23.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/FgQdlzbQ88/media/0.jpg","type":"image","width":2268,"height":2272}],"quotedPost":{"id":"2072648051416826339","platform":"x","url":"https://x.com/every/status/2072648051416826339","shareMode":"full","content":{"text":"Codex works best when the setup matches how you work. Long-running threads, local context folders, outcome-first prompts — our team’s setups look nothing alike. (@tedescau refuses to search for specific files, for example) https://t.co/7yadOEZfKs"},"author":{"name":"Every 📧","handle":"every","avatar":"https://pbs.twimg.com/profile_images/1353773856159195143/Fb-kmsOc_200x200.jpg"},"metadata":{"likes":67,"shares":4,"timestamp":"2026-07-02T11:46:18.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/FgQdlzbQ88/quoted/0.jpg","type":"image","width":960,"height":1200}]},"aiComments":[]},{"id":"2072502518370697311","shareId":"Pvpm0Qd6Ri","platform":"x","url":"https://x.com/XAMTO_AI/status/2072502518370697311","shareMode":"full","createdAt":1783043692710,"postedAt":"2026-07-02T02:08:00.000Z","publishedDate":"2026-07-02T02:08:00.000Z","archivedAt":"2026-07-03T01:54:51.093Z","archivedDate":"2026-07-03T01:54:51.093Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/Pvpm0Qd6Ri","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/BKVzIYvFOe/media/0.png","contentType":"post","content":{"text":"当下AI编程Agent最头疼的问题之一，就是它对你的代码库几乎毫无记忆可言。\n\n每次提问，要么从头到尾重新扫描一遍文件，要么靠向量检索拼凑上下文，Token消耗如流水，结果还不一定精准。codebase-memory-mcp 正是冲着这个痛点来的——给AI装上一套真正持久的代码记忆系统。\n\n这个由DeusData开发的开源项目，将整个代码库索引成一个持久化的知识图谱。它原生支持多达158种编程语言，能够精准提取函数、类定义及调用关系，并通过14个MCP工具将这些结构化信息高效暴露给AI Agent调用。\n\n性能表现亮眼：平均规模的仓库可在毫秒级完成索引，查询响应时间低于亚毫秒级。在代码探索阶段，Token消耗能锐减至原先的二十分之一，大幅降低使用成本。\n\n如果你正使用Claude Code或Cursor处理大型代码库，这个MCP服务器绝对值得一试。\n\n🔗：https://github.com/DeusData/codebase-memory-mcp"},"author":{"name":"Amto","handle":"XAMTO_AI","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/d8546ce0e0dc0123/avatar-9644dceec29c.jpg"},"metadata":{"likes":27,"shares":7,"views":1592,"timestamp":"2026-07-02T02:08:00.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/BKVzIYvFOe/media/0.png","type":"image","width":815,"height":723}],"aiComments":[{"content":"**MCP (Model Context Protocol)**\nAI 시스템과 외부 도구·데이터 소스를 연결하는 오픈 표준 프로토콜. Anthropic이 2024년 11월 공개했으며, AI 에이전트가 파일 읽기·함수 실행·외부 서비스 호출 등을 표준화된 방식으로 수행할 수 있게 한다. [출처](https://en.wikipedia.org/wiki/Model_Context_Protocol)\n\n**LSP (Language Server Protocol)**\n코드 편집기와 언어 분석 서버 간의 통신을 표준화하는 프로토콜. 자동완성, 타입 추론, 심볼 정의 이동 등 언어별 의미론적 분석 기능을 IDE에 제공한다. MCP는 이 LSP의 메시지 흐름 방식을 AI 도구 생태계에 응용한 것이다. [출처](https://modelcontextprotocol.io/)\n\n**AST (Abstract Syntax Tree, 추상 구문 트리)**\n소스 코드의 문법 구조를 트리 형태로 표현한 데이터 구조. 컴파일러나 정적 분석 도구가 코드를 파싱할 때 사용하며, 함수·클래스·호출 관계 등을 기계가 이해할 수 있는 형태로 추출하는 데 활용된다. [출처](https://elbanic.github.io/posts/context-management-with-kiro-part-2/)\n\n**지식 그래프 (Knowledge Graph)**\n엔티티(함수, 클래스 등)와 그 관계(호출 관계, 상속 등)를 노드-엣지 구조로 표현한 그래프 데이터베이스. 코드베이스를 지식 그래프로 인덱싱하면 단순 텍스트 검색보다 구조적·의미론적 질의가 가능해진다. [출처](https://neo4j.com/blog/genai/what-is-model-context-protocol-mcp/)\n\n**벡터 검색 (Vector Search)**\n텍스트나 코드를 수치 벡터(임베딩)로 변환한 뒤 의미적 유사도 기반으로 관련 내용을 검색하는 기법. AI 에이전트가 컨텍스트를 구성할 때 사용하지만, 구조적 정보 없이 텍스트 조각을 짜깁기하는 방식이라 정확도 한계가 있다. [출처](https://neo4j.com/blog/genai/what-is-model-context-protocol-mcp/)","meta":{"id":"ai-action-comment-aiaj_b7967d6e-3ac0-4c08-883a-c3f851490163","cli":"claude","model":"sonnet","type":"glossary","generatedAt":"2026-07-03T07:07:16.904Z"}}]},{"id":"2069010198040326329","shareId":"M1bKQHzydS","platform":"x","url":"https://x.com/hey_madni/status/2069010198040326329","title":"How to Turn Claude Code Into a Full Team of Specialists:","shareMode":"full","createdAt":1783043679637,"postedAt":"2026-06-22T10:50:46.000Z","publishedDate":"2026-06-22T10:50:46.000Z","archivedAt":"2026-07-03T01:54:37.533Z","archivedDate":"2026-07-03T01:54:37.533Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/M1bKQHzydS","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/cMkc7u5fja/media/0.jpg","contentType":"article","content":{"text":"How to Turn Claude Code Into a Full Team of Specialists:\n\nhttp://x.com/i/article/2068971860516315136\n\n---\n\nwhat most people do with Claude Code:\n- open a session\n- paste a task\n- get the thing done\n- close the tab\nNext session you explain it all over again. \nYour stack. Your standards. \nThe way you like things shipped. \nWhat \"done\" even means on your team.\nThat is one brilliant assistant with amnesia.\nUseful. But small.\nThere is a different way to run it.\nClaude Code out of the box is raw. Ridiculously capable, but raw. \nit knows how to code, write, research, edit video. \nWhat it doesn't know is you. It starts cold every single time. No memory of your conventions, your voice, your benchmarks, the checklist that keeps you from shipping something broken at 2am.\nSo you spend the first ten minutes of every session re-teaching it. \nAgain. And again.\nClaudeKit fixes that (http://theclaudekit.com/)\n\n\n## What ClaudeKit actually is\nNot a prompt pack.\nNot a Notion doc full of \"10 killer prompts for developers.\"\nNot a collection of tips you copy and paste and forget by Thursday.\nIt's a marketplace of vertical kits for Claude Code. \nEach kit is a complete specialist. Slash commands, skills, and subagents, all pre-built for one type of work and one type of person.\nYou install a kit, and Claude Code *becomes* that specialist. \nInstantly. \nIt shows up already knowing the workflow, already carrying the domain knowledge, already wired to the tools that job needs.\nYou are not prompting from scratch anymore.\nYou are hiring a department.\nAny kit starts at $14.99 a month. \nInstalling is one line inside Claude Code:\n```\n`ck install [kitname]`\n```\nThat's the whole setup. \nbrowse every kit at https://theclaudekit.com before you spend a cent\nAnd yeah, every file is measured for context cost before you load it so you always know what a kit is doing to your window. \nThat part is kind of the whole point.\nLet me walk you through the team.\n\n## EngineerKit\n![image](https://pbs.twimg.com/media/HLaFKD8b0AA9qnb.jpg)\nThis is the one I reach for daily\n25 commands, plus skills and read-only subagents, all built around one rule: nothing ships on a vibe. \nEverything ends with evidence.\n```\n/eng-debug\n```\npulls the trace, reads the last 15 commits near the failure, reproduces the bug first refuses to touch code until it can name the cause in one sentence ships the smallest fix plus the regression test that would have caught it red before, green after, full suite run.\n```\n/eng-review \n```\nreviews your diff and flags only what would actually block a merge confidence level on every finding not here to nitpick quotes (here to tell you what's going to break)\n```\n/eng-ship \n```\ncommits, pushes, opens the PR with a description written from the actual diff flat out refuses if tests are failing or a secret is sitting in the diff it won't let you ship the broken thing\n```\n/eng-catchup\n```\n is the memory cure. \nCame back after a pull, a /clear, or a long weekend??? It reads what changed so you're not re-explaining your own repo to your own assistant.\nFor devs, solo founders, & agencies shipping real products on real deadlines. \nIf you write code with Claude Code, this is the floor.\n\n## VideoKit\n![image](https://pbs.twimg.com/media/HLaFp6PacAAcC99.jpg)\nI did not expect a video studio to live inside my terminal. \nIt does now...\n17 commands that take video from \"I have an idea\" to \"the file is rendered & it's good.\"\n```\n`/video-make` \n```\ngoes from a description a URL, or even a repo straight to a finished video. Music, transitions, the whole thing. \nAnd it watches its own output afterward a real post-render check so what lands on your disk is confirmed not a maybe that's secretly black frames for 4 seconds.\n```\n/video-clone\n```\nis the one that makes people lean in. \nPoint it at a video you wish you'd made. \nIt measures every scene of the reference. \nThe cuts the palette, the type, the motion and easing. \nThen it rebuilds that exact style in Remotion with your content, and re-analyzes its own render to prove the match. \nScene count, durations, cuts per ten seconds, all compared back to the original. Their pacing and taste, your footage and copy.\n```\n/video-social \n```\ntakes one long video and cuts it for TikTok, Reels, Shorts, LinkedIn, and X in a single run. \nEvery format, one command.\n```\n/video-repurpose \n```\nfinds the best moments in a long piece automatically, so you're not scrubbing a timeline at midnight looking for the good 12 seconds.\nFor content creators, marketers, & agencies who are drowning in \"we need it in every aspect ratio by tomorrow.\"\n\n## MarketingKit\n![image](https://pbs.twimg.com/media/HLaGdIhbAAAv5eA.jpg)\nEvery other kit does marketing ops. This one does the thing nobody else nails. It makes the writing sound like a human wrote it.\n```\n/mkt-humanize\n```\nis the moat. It sweeps a draft for the 14 tells that scream \"AI wrote this.\" Em dashes used as glue. \nThe \"it's not X it's Y\" tic. The banned words, delve and leverage and seamless and elevate, the whole tired vocabulary. It doesn't just flag them. \nIt fixes them, rewrites the piece in your voice, & shows you the worst 5 fixes as before and after lines so you can actually see what changed. \n```\n/mkt-launch \n```\nwrites the full launch set in one pass. \nThe X post, the LinkedIn post, the Product Hunt tagline and description, and the launch email. \nAll of it run through the voice engine, so the launch reads like you wrote it on a good day, not like a template.\n```\n/mkt-competitors\n```\ntears down a competitor's content engine. \nTheir hooks, their cadence, their formats, and the gaps worth stealing.\nAnd before any of that `/mkt-voice` builds a voice file from 5 to 10 of your real posts. \nYour sentence rhythm, your openers, the words you actually own. \nAfter that every command in the kit writes like you instead of like a chatbot doing an impression of you.\nFor founders doing their own marketing, content teams, and agencies tired of editing the AI smell out of every draft by hand.\n\n## SEOKit\n![image](https://pbs.twimg.com/media/HLaG6RGacAAxbxa.jpg)\nSEO is mostly knowing where the free money is hiding. \nThis kit finds it.\n```\n/seo-quick-wins \n```\nis the first thing I run on any site. It pulls 90 days of Search Console data, finds the keywords you almost rank for positions 8 to 20 where Google already half-trusts you, plus the snippets at the top quietly losing clicks. \nThen it ranks every win by click prize, in actual clicks per month, with a one-line fix per row. \n```\n/seo-keywords\n```\ngives you keyword clusters with volume, difficulty, and the part most tools skip: whether the person searching is researching or ready to buy. \nIntent is the whole game.\n```\n/seo-audit \n```\naudits the entire site and hands you one readable report. \nTop 3 problems first, each written as evidence, impact, and the fix. \nNot a 4,000 row CSV dump you'll never open. \nThe actual stuff that's costing you traffic in order.\nFor content teams, SEO agencies& founders growing on organic instead of burning cash on ads.\n\n## EcomKit\n![image](https://pbs.twimg.com/media/HLaHptqbsAAphQw.jpg)\nThis kit talks in money. \nEvery output lands in dollars, hours, or margin points.\n```\n/ecom-no-sales\n```\nis the reality check. It compares your conversion rate to the real benchmark for your price band, not some flat 1.4% average that means nothing. Then it walks your store like an actual shopper would. \nHomepage, collection, best-seller, cart. And it hands back at most 8 fixes ranked by expected dollars per month. If the truth is that your traffic is the problem, it says so to your face instead of blaming your product page.\n```\n/ecom-flows \n```\nwrites all 5 core Klaviyo flows with full email copy in your brand voice, send timing, and discount logic baked in. Welcome, abandoned cart, browse abandon, post-purchase, winback. \n```\n/ecom-ads\n```\nbuilds a 27-ad test batch, 3 hooks by 3 bodies by 3 CTAs, pulled straight from your own customers' review language. \nThe words buyers already use, turned into the words that make them buy.\nFor ecom founders, Shopify operators, and agencies running stores for clients who only care about one number.\n\n## TradeKit\n![image](https://pbs.twimg.com/media/HLaIh34aIAAuHSp.jpg)\nThe bonus seat at the table...\nIt's an AI quant research desk for prediction markets & crypto.\n```\n/tk-scan \n```\nscans live Polymarket and Kalshi markets and ranks them by an edge score, where your model and the crowd's price disagree.\n```\n`/tk-dashboard` \n```\nspins up a full quant dashboard right in your browser, built from your own markets. Live scanner, sizing, the works.\nAnd now the honest part, because this is the bit most tools lie about: it's read-only and paper by default. \nIt does not touch your money. Roughly 84% of Polymarket wallets lose, and any tool promising you profits is selling you something. \nTradeKit is a research desk, not a money printer. \nIt makes you sharper and keeps you disciplined. \nThat's it. \nFor prediction market traders & crypto researchers who want an edge without the fairy tale.\n\n## Getting started\nThree steps. That's the whole thing....\n- Go to theclaudekit.com and look at the kits. \n- Pick the one that matches the work you actually do this week.\n- Run ck install [kitname] & follow the steps\n- And watch it turn into that specialist in front of you.\nThat's it. You're running a team now.\nMost people will read this whole thing, nod, and install nothing. \nThey'll keep opening cold sessions and re-explaining their repo to Claude every single time.\nThe ones who run that first command today will have a full team by the end of the week.\nAnd they are not going back to a single chat box.\nhttps://theclaudekit.com\nGo hire your department.\n— Madni"},"author":{"name":"Madni Aghadi","handle":"hey_madni","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/3103f863cfce0dad/avatar-6c8e1fc3fa17.jpg","bio":"i build stuffs with CC & @theclaudekit"},"metadata":{"likes":221,"shares":33,"views":304526,"timestamp":"2026-06-22T10:50:46.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/cMkc7u5fja/media/0.jpg","type":"image","width":4096,"height":1638},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/cMkc7u5fja/media/1.jpg","type":"image","width":1983,"height":793},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/cMkc7u5fja/media/2.jpg","type":"image","width":1983,"height":793},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/cMkc7u5fja/media/3.jpg","type":"image","width":1983,"height":793},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/cMkc7u5fja/media/4.jpg","type":"image","width":1983,"height":793},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/cMkc7u5fja/media/5.jpg","type":"image","width":1983,"height":793},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/cMkc7u5fja/media/6.jpg","type":"image","width":1983,"height":793}],"aiComments":[]},{"id":"2072691780609589319","shareId":"iAXiaqYmKf","platform":"x","url":"https://x.com/hey_madni/status/2072691780609589319","title":"ANTHROPIC이 FABLE 5를 위한 공식 프롬프팅 가이드를 발표했습니다","shareMode":"full","createdAt":1783043662419,"postedAt":"2026-06-22T10:50:46.000Z","publishedDate":"2026-06-22T10:50:46.000Z","archivedAt":"2026-07-03T01:54:18.984Z","archivedDate":"2026-07-03T01:54:18.984Z","viewCount":2,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/iAXiaqYmKf","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/mgYJJcK3YP/media/0.jpg","contentType":"article","content":{"text":"# ANTHROPIC이 FABLE 5를 위한 공식 프롬프팅 가이드를 발표했습니다\n\n## 1\n\nANTHROPIC이 FABLE 5를 위한 공식 프롬프팅 가이드를 발표했습니다\n\n그들의 문서 속에 묻혀 있습니다\n\n아무도 읽지 않고 있어요\n\n그래서 제가 여러분을 위해 하나의 치트시트로 바꿔봤습니다\n\n대부분의 사람들이 놓치는 가장 큰 점:\n오래된 Opus 4.8 프롬프트는 Fable 5를 적극적으로 더 나쁘게 만듭니다\n\n다음 세션 전에 이걸 저장하세요\n\n참고로 ClaudeKit은 Fable 5를 위한 Claude Code 전문 기술을 제공합니다\n\ntheclaudekit(.)com\n\n전체 가이드는 아래에 있습니다\n\n{{IMAGE_0}}\n\n{{IMAGE_1}}\n\n---\n\n## 2\n\nAnthropic's official Fable 5 prompting guide:\n\nhttps://t.co/fKCAHDhgE9"},"author":{"name":"Madni Aghadi","handle":"hey_madni","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/3103f863cfce0dad/avatar-6c8e1fc3fa17.jpg","bio":"i build stuffs with CC & @theclaudekit"},"metadata":{"timestamp":"2026-06-22T10:50:46.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/mgYJJcK3YP/media/0.jpg","type":"image","width":453,"height":680},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/mgYJJcK3YP/media/1.jpg","type":"image","width":680,"height":272}],"comments":[{"id":"2072694386853597669","content":"awesome. thanks for the link","author":{"name":"Rahul","username":"sairahul1","handle":"sairahul1","url":"https://x.com/sairahul1"},"timestamp":"2026-07-02T14:50:25.000Z","replies":[{"id":"2072695553033740604","content":"glad it helped!","author":{"name":"Madni Aghadi","username":"hey_madni","handle":"hey_madni","url":"https://x.com/hey_madni"},"timestamp":"2026-07-02T14:55:03.000Z"}]},{"id":"2072693595316818242","content":"the fact that old prompts actively brick the new model is so easily overlooked\n\nthis guide will save ppl hours","author":{"name":"Melio","username":"MelioHL","handle":"MelioHL","url":"https://x.com/MelioHL"},"timestamp":"2026-07-02T14:47:16.000Z","replies":[{"id":"2072693886955442634","content":"hopefully","author":{"name":"Madni Aghadi","username":"hey_madni","handle":"hey_madni","url":"https://x.com/hey_madni"},"timestamp":"2026-07-02T14:48:26.000Z"}]},{"id":"2072821068692414493","content":"Helpful Guide","author":{"name":"Adarsh Chetan","username":"AdarshChetan","handle":"AdarshChetan","url":"https://x.com/AdarshChetan"},"timestamp":"2026-07-02T23:13:48.000Z"},{"id":"2072713484228288567","content":"reading the official docs is cool but a cheatsheet saves actual time\n\nwhats the biggest change they recommend?","author":{"name":"Amine","username":"amine_8_8","handle":"amine_8_8","url":"https://x.com/amine_8_8"},"timestamp":"2026-07-02T16:06:18.000Z"},{"id":"2072697284769300728","content":"@hey_madni\n bro, didn't realize old prompts could mess things up. good heads-up, I've been using the same stuff from Opus 4.8 without even thinking about it","author":{"name":"Hussain Hashim | Building SundayBack","username":"itsthedonhashim","handle":"itsthedonhashim","url":"https://x.com/itsthedonhashim"},"timestamp":"2026-07-02T15:01:56.000Z"},{"id":"2072842792196297048","content":"I don’t understand how to create sub agents","author":{"name":"305office","username":"305office","handle":"305office","url":"https://x.com/305office"},"timestamp":"2026-07-03T00:40:07.000Z"},{"id":"2072736017669570853","content":"cool,\nRather than reading the whole guide this is helpful.\nThanks for the sharing.","author":{"name":"Rohittt","username":"iRohit21","handle":"iRohit21","url":"https://x.com/iRohit21"},"timestamp":"2026-07-02T17:35:50.000Z"},{"id":"2072694139045499008","content":"If AI isn't solving a real problem for you, you're using it wrong.","author":{"name":"Ifeoluwapo Taiwo","username":"Ingeniousife","handle":"Ingeniousife","url":"https://x.com/Ingeniousife"},"timestamp":"2026-07-02T14:49:26.000Z"}],"aiComments":[]},{"id":"2072698394254995602","shareId":"dGWzv2r09V","platform":"x","url":"https://x.com/0xCodez/status/2072698394254995602","shareMode":"full","createdAt":1783043647718,"postedAt":"2026-07-02T15:06:20.000Z","publishedDate":"2026-07-02T15:06:20.000Z","archivedAt":"2026-07-03T01:54:06.245Z","archivedDate":"2026-07-03T01:54:06.245Z","viewCount":6,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/dGWzv2r09V","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072698092340629505/img/KlsFsqJTDsKNODKb.jpg","contentType":"post","content":{"text":"Creator of OpenClaw: \n\n\"in January, juggling 10+ terminals felt like peak productivity. Today it feels silly. \n\ni thought I was orchestrating - really, I was the scheduler, the router, and the memory\" \n\nin this 6-minute talk, the founder of OpenClaw shares how he builds effective agentic loops from scratch.\n\nWatch it today, then explore how to build loops for trading agents in artice below"},"author":{"name":"Codez","handle":"0xCodez","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/62bc1cc8bb023f27/avatar-dfccb1abb8db.jpg"},"metadata":{"likes":107,"shares":14,"views":9313,"timestamp":"2026-07-02T15:06:20.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2072698092340629505/vid/avc1/1280x720/zfFymoZGFBcUhPrF.mp4?tag=28","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072698092340629505/img/KlsFsqJTDsKNODKb.jpg","type":"video","width":1920,"height":1080,"duration":373}],"quotedPost":{"id":"2072668933690126571","platform":"x","url":"https://x.com/0xMovez/status/2072668933690126571","shareMode":"full","content":{"text":"Loop Engineering for trading agents: 12 Steps to build a self-running Quant desk\n\nhttp://x.com/i/article/2071890135302426625"},"author":{"name":"Movez","handle":"0xMovez","avatar":"https://pbs.twimg.com/profile_images/1998148360478322688/851J4fBL_200x200.jpg"},"metadata":{"likes":190,"shares":33,"timestamp":"2026-07-02T13:09:16.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/FWb4FGaInw/quoted/0.jpg","type":"image","width":1200,"height":480},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/FWb4FGaInw/quoted/1.png","type":"image","width":688,"height":268},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/FWb4FGaInw/quoted/2.png","type":"image","width":868,"height":774},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/FWb4FGaInw/quoted/3.jpg","type":"image","width":1200,"height":667}]},"aiComments":[]},{"id":"2072668933690126571","shareId":"5YUSxKQZLa","platform":"x","url":"https://x.com/0xMovez/status/2072668933690126571","title":"Loop Engineering for trading agents: 12 Steps to build a self-running Quant desk","shareMode":"full","createdAt":1783043635444,"postedAt":"2026-07-02T13:09:16.000Z","publishedDate":"2026-07-02T13:09:16.000Z","archivedAt":"2026-07-03T01:53:53.467Z","archivedDate":"2026-07-03T01:53:53.467Z","viewCount":7,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/5YUSxKQZLa","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/0.jpg","contentType":"article","content":{"text":"Loop Engineering for trading agents: 12 Steps to build a self-running Quant desk\n\nhttp://x.com/i/article/2071890135302426625\n\n---\n\nThe leverage in investing moved. It’s no longer about watching charts faster - it’s about designing the loop that researches, decides, and trades for you. \nHere are the 12 steps, end to end, inside one app.\nMost traders still run themselves like a script. They open ten tabs, read the same charts, type the same questions, and close the laptop - only to start over tomorrow.\n9 in 10 never build a single loop that does the watching for them. No research that runs on a schedule, no strategy that backtests itself, no execution that fires while they sleep.\n> Follow my Linkedin for AI alpha: https://linkedin.com/in/lev-deviatkin\nThis is the 12-step roadmap to a quant desk that runs itself, built entirely inside Minara.\nEvery loop has the same five-beat anatomy. Hold this shape in your head - every one of the 12 steps below slots into one of these beats:\n![image](https://pbs.twimg.com/media/HMDWe1rWEAAgi-G.png)\n[Minara](https://minara.ai/app/chat?r=0XMOVEZ) closes that loop in one app across three surfaces, each owning a different stretch of the cycle. The whole guide is organized around them.\nYou don’t need all three on day one. Most desks start with a single scheduled research Workflow and grow from there. \nBut the power compounds when the surfaces connect - the output of one becomes the input of the next, which is the whole point of a loop.\n![image](https://pbs.twimg.com/media/HMDW8sTW8AAW_vI.png)\n12 steps. 3 loops. One desk that runs without you.\n\n\nLoop 1 • The Research Loop\n\n## 01. Ask the market a real question\nEvery loop starts with a question worth answering on a schedule. \nIn Chat, you don’t search - you ask, and Minara pulls live market, on-chain, sentiment, and macro data behind the answer. \nOpen the[Minara AI](https://minara.ai/app/chat?r=0XMOVEZ)chat and enter the following test command:\n```python\nHow is the bond market signaling risk appetite for \nstocks and crypto right now?\n\n```\n![image](https://pbs.twimg.com/media/HMDZD9lW8AAHWfr.png)\nThe answer isn’t a wall of links it’s a structured read with the data behind it. That structure is what makes it loopable: a question that returns a consistent shape every time is a question you can schedule later (step 9).\n![image](https://pbs.twimg.com/media/HMDZPx5WgAApYSg.png)\nThree more opening questions that produce schedulable answers - the difference is they each return a repeatable structure, not a one-off opinion:\n```python\nWhat market phase are we in right now — risk-on, risk-off, \nor transition? Give me the three signals that matter most \nand how I should be positioned.\n\n```\n```python\nWhat are the top 5 Polymarket leaderboard traders betting on \nthis week, and is there a consensus worth fading?\n\n```\nBefore you ask anything, run this filter: would this question still be useful if it ran by itself every morning? “\n\n\n\n## 02. Deep-dive a single asset\nOnce the macro read points somewhere, narrow to one asset and pull the full picture - price action, on-chain flows, derivatives positioning, sentiment. This is the work that normally costs you five tabs and an hour.\n```python\nDeep-dive $HYPE: funding rates, open interest, whale moves,\nand sentiment over the last 7 days. Where’s the risk?\n\n```\n![image](https://pbs.twimg.com/media/HMDaXgAWsAAbibh.png)\n![image](https://pbs.twimg.com/media/HMDagq5XgAANbV1.png)\nThe reason this works for a loop is that [Minara](https://minara.ai/r/28M511) reads across six data domains in a single pass. Knowing what’s available tells you what to ask for - here’s the map:\n- Markets: live prices, OHLCV, ETF/macro, financial analytics\n- Onchain: wallet assets, transfers, top holders, unlocks, whale alerts\n- Signals: whale moves, news, sentiment, narrative signals, calendars\n- Derivatives: funding rates, OI, liquidations, exchange flows, borrow rates \n- Predictions: Polymarket search, positions, leaderboards, activity\n- DeFi: pool data, yield discovery, DEX trends, Pendle markets\nThe deep-dive prompt that consistently produces a tradeable read names the domains explicitly - it tells Minara which lenses to combine rather than leaving it to guess:\n```python\nFull thesis on $SOL: trend (markets), holder concentration (onchain), \nfunding + OI skew (derivatives), and narrative momentum (signals). \nEnd with a one-line risk/reward call.\n\n```\n\n\n\n## 03. Audit your own trades\nA self-improving loop needs a verifier - something that grades the work instead of agreeing with it. Before you automate anything, point Minara at your own track record. \nThe honest read on what you keep getting wrong is the most valuable input to every loop that follows.\n```python\nReview my trades from the last month and suggest improvements. \nWhere am I entering too early or sizing too big?\n\n```\nThe model that writes a strategy is too easy on it. A separate pass that only critiques - your trades, your sizing, your timing - is the cheapest verifier you have.\n\n\n\n## 04. Turn the hunch into a thesis\nThe last research step converts everything above into a written thesis with explicit entries, exits, and risk - the spec a strategy will be built from. \nMake [Minara](https://minara.ai/app/strategy-studio?r=0XMOVEZ) commit to numbers, not vibes.\n![image](https://pbs.twimg.com/media/HMDb_dPX0AAFPzm.png)\nThe difference between a thesis you can build a strategy from and one you can’t comes down to specificity. \nA loop can only automate what’s measurable - so the thesis has to be measurable too:\n![image](https://pbs.twimg.com/media/HMDcLJ-WkAAtwow.png)\nResearch that returns the same shape every time is research you can schedule.\n\n\nLoop 2 • The Strategy Loop\n\n## 05. Prompt the strategy in plain language\n[Strategy Studio](https://minara.ai/app/xstrategy?r=0XMOVEZ) reads a sentence and emits a structured strategy spec -entries, exits, sizing, universe. No code, no quant background. \nThe thesis from step 4 becomes a working strategy you can test in the same thread. These are the real prompt patterns the Studio is built around:\n```python\n$HYPE momentum strategy (15m): long on ROC>0 + volume breakout, \nhold 10 bars; 10x, stop 2% target 5%, 6-month backtest.\n\n```\n![image](https://pbs.twimg.com/media/HMDddwaWIAAE-ah.png)\n![image](https://pbs.twimg.com/media/HMDdq0wWUAIof0g.jpg)\n> Get a Lite plan & 1400 credits for free on Minara with my link: https://minara.ai/app/strategy-studio?r=0XMOVEZ ( limited 200 users )\n\nTwo structural choices shape what you build. Get them right and the backtest means something, get them wrong and you’re testing the wrong thing entirely:\n- Time-series: One asset through time. “When does this coin trigger?” Momentum, mean-reversion, breakout - anything keyed to an asset’s own history.\n- Cross-sectional: Many assets at one moment. “Which coins rank top today?” Relative strength, rotation, pairs - anything that compares a universe.\n![image](https://pbs.twimg.com/media/HMDeQrIWEAAHFQq.png)\nAnd you don’t have to start from a sentence. Studio takes a thesis in four forms - pick whichever matches what you already have:\n- Plain language - describe entries, exits, sizing in a sentence (the prompts above).\n- Build with a Form - fill in asset, timeframe, and preferences in structured fields when you want precision over prose.\n- Video to Strategy - saw a strategy breakdown on YouTube? Upload it and Studio generates a ready-to-trade version.\n- Code to Strategy - paste Pine Script or code from any platform and it ports across.\n\n\n\n## 06. Start from a proven template\nIf you’d rather not write the thesis from scratch, Studio ships four template families - each a known-good baseline you can fork and tune:\n![image](https://pbs.twimg.com/media/HMDenrsXgAAJ0mm.jpg)\n- Momentum - ride persistent trends with adaptive stops and volatility-aware sizing.\n- Mean-reversion - fade extreme moves around a moving anchor when volume confirms exhaustion.\n- Arbitrage - capture price dislocations across venues, pairs, and funding rates in real time.\n- Pairs - long one asset, short its cointegrated twin; neutral to the tape, sensitive to the spread.\nPick the family that matches your thesis, then describe the asset and timeframe. The Studio handles the structure; you bring the edge.\n\n\n\n## 07. Backtest with risk built in\n[Studio](https://minara.ai/app/xstrategy?r=0XMOVEZ) replays the strategy against 10+ years of market data in seconds -modeling fees, funding, borrow, and slippage from venue-specific cost curves. \nYou get the equity curve, but more importantly you get the risk beside the return: rolling max-drawdown, a forward volatility cone, and an exposure map so concentration never hides.\n![image](https://pbs.twimg.com/media/HMDe795XEAAtxW0.png)\nWalk-forward, out-of-sample, and regime-sliced views come on every run, and a leakage check flags suspicious results automatically - the difference between a backtest that survives live and one that only looked good on paper.\nThe metric that decides whether a strategy is worth deploying isn’t total return - it’s return per unit of risk. Two strategies can post similar returns with completely different risk profiles.\n[Minara](https://minara.ai/app/strategy-studio?r=0XMOVEZ) puts the risk right next to the reward, three ways:\n- Drawdown - the worst path, visible.\n![image](https://pbs.twimg.com/media/HMDnYhJWsAAAxjT.jpg)\n- Volatility cone - calibrated to regime\n![image](https://pbs.twimg.com/media/HMDff2FXkAAHtwl.png)\nAnd the exposure map so concentration risk never hides - a per-symbol view of where the capital actually sits. Darker means heavier weight:\n![image](https://pbs.twimg.com/media/HMDfl-4WsAAxnLA.png)\nA finished backtest card - e.g. HYPE 15m Leveraged MACD, +366.15%, with the equity curve and the Max Drawdown / Win Rate / Sharpe row underneath.\n![image](https://pbs.twimg.com/media/HMDfx1VXgAAEEi5.png)\n\n\n\n## 08. Paper trade before real capital\nA backtest is a promise. Paper trading is the receipt. Promote any backtest to paper with one click and Studio runs it on the same engine that powers live execution - same fee model, same slippage curve, same risk hooks -against live market data, no capital at stake.\nWhen the paper run confirms what the backtest promised, the strategy has earned the right to real money.\nWhen it doesn’t, you learned that for free. This is the verifier gate of the strategy loop - nothing goes live until the paper run says yes.\n![image](https://pbs.twimg.com/media/HMDgOwiXkAAXdXR.png)\nBefore you promote a paper run to live, it should clear five checks. This is the gate that separates a strategy that survives from one that only looked good in a replay:\n- Paper P&L tracks the backtest - within a reasonable band, not wildly below. A big gap means the backtest was optimistic about fills.\n- The leakage check is clean - no future data crept into the signal.\n- Out-of-sample holds up - the strategy works on data it wasn’t tuned on, not just the in-sample window.\n- Drawdown is survivable - you could sit through the worst path shown without pulling the plug.\n- It works across regimes - the regime-sliced view doesn’t show all the profit coming from one lucky month.\nA backtest is a promise. Paper trading is the receipt.\n\n\nLoop 3 • The Execution Loop\n\n## 09. Schedule the watching with Workflow\nThis is where the desk stops waiting for you. [Workflow](https://minara.ai/app/workflow?r=0XMOVEZ) is no-code automation in plain language - you describe what should happen and when, and Minara builds the monitor.\n The example prompts the product ships with show exactly the shape:\n```python\n30 minutes before the U.S. market opens, select 4 stocks and email \nme their tickers, entry prices, and a brief research note.\n\n```\n```python\nSet up a weekly market report to my email — tell me what phase we’re \nin, why it matters, and how I should adjust my strategy.\n\n```\n![image](https://pbs.twimg.com/media/HMDhcvXXgAAnil6.png)\nTwo more example prompts the product ships with - both produce a recurring monitor that emails you on a schedule:\n```python\nSend me daily reports of the moves and on-chain activity of the top 3 \nholdings in my Minara wallet.\n\n```\n```python\nEvery Sunday evening, email me a summary of the week’s major macro \nevents and how they affected the market.\n\n```\nIf you’d rather not write a prompt, Quick Templates build standard workflows from simple form fields - Polymarket Address Monitor, Polymarket Odds Monitor, and Copy Trade are ready out of the box.\n Fill the fields, deploy, done.\n\n\n\n## 10. Start alerts-only, then let it act\nThe safest loop earns trust before it earns permission. \nEvery [Workflow](https://minara.ai/app/workflow?r=0XMOVEZ) can run alerts-only first - it watches and notifies, but never executes. You read its calls for a week. When its judgment matches yours, you let it act.\n```python\nBuy 200 USDT of $SOL if price ≤ 175 USDT, then take profit at 200 \nUSDT and stop loss at 160 USDT.\n\n```\nThat single sentence becomes a conditional order routed through Minara’s monitor system for high-precision execution. Keep Telegram or email notifications on and you get a real-time receipt for every fire.\n\n\n\n## 11. Hand the rules to Autopilot\nA proven strategy goes live with one click to Autopilot, which runs it on Hyperliquid with deterministic rules - no discretionary trades, no hidden heuristics. \nWhen an entry or exit rule fires (a Supertrend flip, an RSI threshold, a grid level), Autopilot acts. \nEvery position ships with mandatory take-profit and stop-loss, and the stop trails as the trade moves.\n![image](https://pbs.twimg.com/media/HMDinV-XgAA-ArQ.png)\nYou stay in command of the boundaries: you authorize exactly which assets Autopilot can touch, set the leverage scope per symbol, and set an Initial Equity Drawdown Limit that flattens everything if hit. \nClose a position, pause the engine, or pull an asset from scope at any time - every manual action is treated as an intentional override, with no hidden retries.\nYou can start from an official strategy instead of building your own. Each comes with its own preset trading scope and risk framework:\n- Sharpe Guard: 15m trend-following\n- Supertrend Monitor: Multi-timeframe\n- Classic Futures Grid: Range-bound \n- Custom (Studio): Your\nWhatever strategy runs, four risk controls are non-negotiable - they’re what makes handing over the wheel a discipline rather than a gamble:\n- Mandatory TP/SL - every position opens with take-profit and stop-loss attached; they can’t be silently removed while Autopilot is active.\n- Trailing stop - the stop moves with the trade to lock in gains; if technicals turn sharply, Autopilot can close at market rather than wait.\n- Equity drawdown limit - set a hard floor; if account equity hits it, everything flattens. The trade-level stop protects the trade; this protects the account.\n- Trading scope - Autopilot only touches assets you authorize, at the leverage you allow. Everything outside scope stays in your hands.\n\n\n\n## 12. Close the loop - and let it compound\nHere’s where twelve steps become one system. \n![image](https://pbs.twimg.com/media/HMDjuNuW4AA5mwt.png)\n- The Workflow that runs your morning research (step 9) feeds the thesis you refine in Chat (steps 1–4).\n- That thesis becomes a strategy in Studio (steps 5–8). The strategy goes live on Autopilot (step 11). \n- Autopilot’s results become next week’s trade review (step 3) - and the loop starts again, sharper.\nNotice the three verifier gates in the middle - the trade audit (step 3), the paper run (step 8), and alerts-only mode (step 10). They’re what keep an automated desk honest. \nA loop with no gate is just an agent agreeing with itself at speed, a loop with gates is a system that catches its own mistakes before they cost you.\nOne Workflow keeps it honest: a weekly review that grades the desk against itself\n```python\nEvery Sunday evening, email me a summary of the week’s major macro \nevents, how they affected my open positions, and what I should \nadjust.\n\n```\nThat’s the whole discipline: research feeds strategy, strategy feeds execution, execution feeds review, review feeds next week’s research. You designed it once. Now it runs - and every loop leaves the next one a little smarter.\n\n\n\n## Conclusion: \nStop running yourself like a script. Design the loop instead.\nFor two years, trading edge meant reading faster, refreshing more tabs, catching the move before the next person. That phase is ending. \nThe edge now is the loop you design - the research that runs on a schedule, the strategy that backtests itself, the execution that fires on rules while you’re asleep.\n[Minara](https://minara.ai/app/strategy-studio?r=0XMOVEZ) closes that loop in one app: ask in Chat, build in Strategy Studio, automate in Workflow, execute on Autopilot. Twelve steps from a hunch to a desk that runs itself - no code, full control, real risk discipline at every gate.\nPick one step you’re not doing - probably your first scheduled Workflow, or a paper run before going live - and set it up today. Then the next. The desk compounds from there."},"author":{"name":"Movez","handle":"0xMovez","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/1c08021a94665113/avatar-5cf839236e7c.jpg","bio":"Content creator | AI researcher & agentic builder | @zscdao"},"metadata":{"likes":190,"shares":33,"views":241289,"timestamp":"2026-07-02T13:09:16.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/0.jpg","type":"image","width":1983,"height":793},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/1.png","type":"image","width":688,"height":268},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/2.png","type":"image","width":868,"height":774},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/3.jpg","type":"image","width":1409,"height":783},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/4.png","type":"image","width":685,"height":265},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/5.jpg","type":"image","width":912,"height":320},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/6.png","type":"image","width":685,"height":155},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/7.png","type":"image","width":687,"height":361},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/8.png","type":"image","width":724,"height":234},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/9.png","type":"image","width":1065,"height":824},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/10.png","type":"image","width":693,"height":139},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/11.png","type":"image","width":670,"height":360},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/12.png","type":"image","width":961,"height":879},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/13.png","type":"image","width":508,"height":423},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/14.png","type":"image","width":851,"height":300},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/15.jpg","type":"image","width":1105,"height":759},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/16.png","type":"image","width":909,"height":662},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/17.png","type":"image","width":692,"height":173},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/18.png","type":"image","width":708,"height":606},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/s0h3vs6dva/media/19.png","type":"image","width":1433,"height":1156}],"aiComments":[]},{"id":"2072728896827482158","shareId":"bhnQ434Olf","platform":"x","url":"https://x.com/AnatoliKopadze/status/2072728896827482158","shareMode":"full","createdAt":1783043631307,"postedAt":"2026-07-02T17:07:33.000Z","publishedDate":"2026-07-02T17:07:33.000Z","archivedAt":"2026-07-03T01:53:48.741Z","archivedDate":"2026-07-03T01:53:48.741Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/bhnQ434Olf","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072727457971527681/img/WS8frtAzvMdOyQx7.jpg","contentType":"post","content":{"text":"OpenAI President, Greg Brockman:\n\n\"Out of a billion ChatGPT users, only about 1% are using agents. \n\nImagine the edge people using agents have right now.\"\n\nInside OpenAI you can't work without an agent anymore, the same way you can't work without Slack.\n\nIn this 40-minute session, he explains everything you have to know about Agents.\n\nWatch it, then read the step by step guide on building loops for your agents below."},"author":{"name":"Anatoli Kopadze","handle":"AnatoliKopadze","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/c079e8c75c83bb64/avatar-5f7419344e40.jpg","bio":"Adapt or die"},"metadata":{"likes":92,"shares":12,"views":13438,"timestamp":"2026-07-02T17:07:33.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2072727457971527681/vid/avc1/1280x720/gTIeLMagY033fUJZ.mp4?tag=28","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072727457971527681/img/WS8frtAzvMdOyQx7.jpg","type":"video","width":1920,"height":1080,"duration":2655}],"quotedPost":{"id":"2072668933690126571","platform":"x","url":"https://x.com/0xMovez/status/2072668933690126571","shareMode":"full","content":{"text":"Loop Engineering for trading agents: 12 Steps to build a self-running Quant desk\n\nhttp://x.com/i/article/2071890135302426625"},"author":{"name":"Movez","handle":"0xMovez","avatar":"https://pbs.twimg.com/profile_images/1998148360478322688/851J4fBL_200x200.jpg"},"metadata":{"likes":190,"shares":33,"timestamp":"2026-07-02T13:09:16.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ya8YG7Q2pQ/quoted/0.jpg","type":"image","width":1200,"height":480},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ya8YG7Q2pQ/quoted/1.png","type":"image","width":688,"height":268},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ya8YG7Q2pQ/quoted/2.png","type":"image","width":868,"height":774},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ya8YG7Q2pQ/quoted/3.jpg","type":"image","width":1200,"height":667}]},"aiComments":[]},{"id":"2072251637742137799","shareId":"Tdz9OxSguz","platform":"x","url":"https://x.com/humzaakhalid/status/2072251637742137799","title":"How to Build $1B/ 1 Person Company Using AI","shareMode":"full","createdAt":1783043609282,"postedAt":"2026-07-01T09:31:05.000Z","publishedDate":"2026-07-01T09:31:05.000Z","archivedAt":"2026-07-03T01:53:27.310Z","archivedDate":"2026-07-03T01:53:27.310Z","viewCount":2,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/Tdz9OxSguz","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/qfOJoMkPcI/media/0.jpg","contentType":"article","content":{"text":"How to Build $1B/ 1 Person Company Using AI\n\nhttp://x.com/i/article/2071895674711658496\n\n---\n\nZero employees. Zero burnout. Real system.\nMost people think a one-person company means doing everything yourself.\nIt doesn't.\nIt means being the only human on the team.\nThe first version of it is 80-hour weeks and a quiet breakdown wearing a business hat. The second version is what I run right now. One person. Three products. Zero employees. The rest is delegated to a Claude Project that knows my business better than most employees ever would.\nBy the end of this issue, you'll get the full One-Person Company OS - every .md file template, every prompt, the exact Claude Project setup I use to run AI in Public. No fluff. No theory. Copy, paste, ship.\n\nLast month I almost shut everything down.\nI was writing newsletters, replying to sponsors, building products, fixing the site, and doing my taxes. \nAll in the same week. I sat down on a Sunday night, and I couldn't even open my laptop. \nThat was the moment I knew. The work wasn't the problem. The setup was.\nI rebuilt everything that week. What you're about to read is what came out of it.\n\n\n## The proof, in case you needed it\nSam Altman said it out loud last year:\n> \"We're going to see 10-person billion-dollar companies pretty soon. In my little group chat with CEO friends, there's this one-person billion-dollar company, which would have been unimaginable without AI.\"\n\nA one-person company will hit a billion dollars in revenue.\nNot a lean startup. Not a small team. One person.\nThe reason it's possible now: AI handles the execution layer. Research, writing, ops, content, customer replies - all of it. \nYou stay in the only seat AI can't fill yet.\nThe decision seat.\nThat's the company we're building here.\n\n\n## Old way vs new way\nOld way: You sit down to write. You start from scratch. You bounce between four tabs, trying to remember what you decided last week. An hour passes. You have one paragraph. You feel behind before you've even started.\nNew way: You open a Project that already knows your voice, your audience, your goals, and last week's decisions. You give it a direction. It hands you a 90% draft in four minutes. You spend the next hour editing - not producing.\nOld way: Every new task = full context dump. \"Here's who I am, here's my audience, here's the tone...\"\nNew way: The Project already has all of that. You say what you need. The context lives in the system. Not in your head.\nThat shift is the whole game.\n![image](https://pbs.twimg.com/media/HMDt8q6XsAAILgx.jpg)\nIf this is useful, share it with one person who's been trying to run everything alone. Takes 10 seconds.\n\n\n## Step 1: Build your Claude Project as a business brain\nA regular Claude chat forgets you every time you close it. A Project doesn't. A Project is a folder where you can drop files, set system instructions, and have every chat inside it remember everything.\nThat's the foundation. Without it, nothing else works.\nQuick setup:\n1. Go to claude.ai → Projects (left sidebar) → New Project\n1. Name it your actual business name. Not \"My Newsletter.\" Not \"AI Stuff.\" The brand.\n1. Add a one-line description: \"Operating system for [business name].\"\n\nThis Project is your operating system. Every chat about your business lives here. Every file lives here. Every decision lives here.\n\"Stop opening a fresh chat for every task. That's how you stay the bottleneck.\"\n\n\n## Step 2: Load the four files that run everything\nEvery one-person company needs exactly four files inside the Project. These are the brain. Without them, Claude gives you generic output. With them, it gives you your output.\nThe four files:\n1. who-i-am.md - your voice and standards\n1. what-i-do.md - your offers and current focus\n1. style-rules.md - the rules that protect your voice\n1. operating-context.md - the running log of your business\n![image](https://pbs.twimg.com/media/HMDwiL0XoAEsW0y.jpg)\n\nHere's the full template for each one. Copy them into Notion or a Google Doc, edit once, then upload to your Project.\nTemplate 1: who-i-am.md\n```\n# Who I Am\n\n## Brand name\n[Your brand name]\n\n## What I do in one line\n[Your one-line pitch. Be specific. Not \"I help people with AI.\" Try\n\"I help solo founders ship content with Claude.\"]\n\n## Who I serve\n- Primary audience: [Who you write for. Be specific. Age, role, problem.]\n- Secondary audience: [The wider crowd that follows along.]\n- Who I do NOT serve: [The wrong-fit reader. Naming this protects your voice.]\n\n## My voice — three words\n[Pick three. Examples: honest, sharp, warm. Or: blunt, useful, calm.]\n\n## What good work looks like\n- A reader finishes and feels they can actually do the thing\n- It sounds like me talking, not a generic AI blog post\n- Every claim is backed by what I actually did, not theory\n- It earns trust by telling readers what didn't work too\n\n## What I will never publish\n- Hype words: \"game-changer\", \"revolutionary\", \"mind-blowing\"\n- Vague advice: \"be more consistent\", \"find your why\"\n- AI-sounding sentences: \"In today's fast-paced world...\"\n- Anything I haven't tested myself\n\n## My non-negotiables\n- Honesty over hype, every single time\n- Specifics over vibes\n- One clear takeaway per piece\n- Reader leaves with something they can use today\n```\n\nTemplate 2: what-i-do.md\n```\n# What I Do\n\n## My products and offers\n1. [Product/offer name] — [one line on what it is and who it's for]\n2. [Product/offer name] — [one line]\n3. [Product/offer name] — [one line]\n\n## Current quarter goals (update every 90 days)\n- Revenue goal: [Specific number]\n- Audience goal: [Specific subscriber, follower, or customer count]\n- Build goal: [The one big thing I'm shipping this quarter]\n\n## What I'm saying YES to this quarter\n- [Specific yes 1]\n- [Specific yes 2]\n- [Specific yes 3]\n\n## What I'm saying NO to this quarter\n- [Specific no 1 — be honest about the shiny things you keep falling for]\n- [Specific no 2]\n- [Specific no 3]\n\n## The one thing that matters most\n[If only one thing got done this quarter, what would it be? Write it\nin one sentence. Read it every Monday.]\n```\n\"The 'NO' list does more work than the 'YES' list. Most people skip it. That's why they stay stuck.\"\n\nTemplate 3: style-rules.md\n```\n# Style Rules\n\n## Words I never use\n- \"Game-changer\", \"revolutionary\", \"unprecedented\"\n- \"In today's fast-paced world\"\n- \"Leverage\", \"synergy\", \"utilize\" (just say \"use\")\n- \"Delve\", \"embark\", \"navigate the landscape\"\n- [Add your own banned words as you spot them]\n\n## Sentence patterns that sound like AI\n- Starting with \"In this article, we'll cover...\"\n- \"It's important to note that...\"\n- \"Whether you're a beginner or an expert...\"\n- Three-item lists where the third item is always a metaphor\n\n## Formatting rules\n- Short sentences. If it can be split, split it.\n- Active verbs. \"Claude builds\" — not \"Claude can build.\"\n- One idea per paragraph.\n- Use arrows (→) for flow and contrast.\n- Bullets for scannable lists, not for hiding from real prose.\n- Em dashes for asides, not hyphens.\n\n## Length rules\n- Newsletter posts: 1,500 to 2,500 words\n- X posts: under 280 characters or threaded\n- LinkedIn posts: 150 to 300 words\n- Reply emails: under 100 words unless it's a relationship reply\n\n## Tone rules\n- Mentor, not teacher\n- Honest, not harsh\n- Specific, not vague\n- Confident, not arrogant\n\n## The final check\nBefore anything goes live, ask: \"Could this have been written by anyone?\"\nIf yes, rewrite until the answer is no.\n```\n\nTemplate 4: operating-context.md\n```\n# Operating Context\n\n## This week (updated every Monday)\n- Focus: [The one thing this week is for]\n- Open loops: [Anything half-done that needs to land]\n- Decisions made: [New decisions made this week that change how we work]\n\n## Active projects\n- [Project 1] — status, deadline, blocker\n- [Project 2] — status, deadline, blocker\n- [Project 3] — status, deadline, blocker\n\n## Recent decisions log\n- [Date]: Decided [X] because [Y]\n- [Date]: Decided [X] because [Y]\n\n## Things I keep changing my mind on (red flag list)\n- [If you flip-flop on something three times, that's a real decision\n waiting to happen.]\n\n## Wins worth remembering\n- [Anything that worked. Anything readers loved. Anything a sponsor said\n yes to. Keep the receipts so future-you can replay them.]\n```\nUpdate this file at the start of every week. Five minutes max. That's all it takes.\n\"This file is the single most important one. It's what stops Claude from giving you last month's advice this week.\"\n\n\n## Step 3: Write your weekly brief, not your weekly prompts\nMost people treat Claude like a vending machine. Put in a prompt. Get out a result.\nThat's not a business system. That's a one-off.\nEvery Monday morning, instead of writing 12 separate prompts across the week, write one brief. Drop it into your Project. Now Claude has your week.\n```\nTemplate: Weekly brief\n# Weekly brief — [Date]\n\n## What's happening this week\n[1 to 3 sentences. What's the focus? What changed since last week?]\n\n## What needs to ship\n- [Deliverable 1 — with deadline]\n- [Deliverable 2 — with deadline]\n- [Deliverable 3 — with deadline]\n- [Deliverable 4 — with deadline]\n- [Deliverable 5 — with deadline]\n\n## What I'm NOT doing this week\n[List the temptations you're saying no to. This trains Claude to\nredirect you when you start drifting.]\n\n## Current state\n[What's changed since last week? New subscribers? A new sponsor?\nA new project? Anything Claude needs to know to be useful this week.]\n\n## The one thing\nIf I only get one thing done this week, it must be: [The one thing.]\n```\n\nIt doesn't need to guess. It doesn't fill gaps with assumptions. It works from what you actually need this week.\n\"The brief is the difference between a tool and a teammate.\"\n\n\n## Step 4: Assign Claude a role, not a task\nThis is the shift that changed everything for me.\nThe biggest mistake I made early on was assigning tasks. \"Write a LinkedIn post.\" \"Draft an email.\" \"Summarize this.\" Tasks produce output. Roles produce systems.\nWhen I gave Claude a role - \"you are the operating partner for AI in Public\" - the quality of everything jumped. It stopped just answering questions. It started thinking about the bigger picture.\n[VISUAL 2: \"Task → Role → System\" - Vertical flow showing the leverage ladder from one-off task to durable system, with a leverage gauge on the left]\nThe role assignment prompt\nDrop this into your Project once. Then never write it again.\n```\nYou are the operating partner for [your business name].\n\nYou have full access to my business files in this Project:\n- who-i-am.md (my voice and standards)\n- what-i-do.md (my offers and quarter goals)\n- style-rules.md (the rules that protect my voice)\n- operating-context.md (the running log of my business)\n\nYour job is not to answer questions on demand. Your job is to help me\nmake better decisions, produce better work, and build systems that\ndon't require me to be present for every task.\n\nThree rules for how you work with me:\n\n1. When I bring you a task, your first question should always be:\n \"Is this something I should do once, or should we build a system\n for it?\"\n\n2. Before producing any content, read style-rules.md and check your\n draft against every rule before you show it to me.\n\n3. If I ask you something that contradicts what's in my files, push\n back. Tell me what the file says. Ask me if I want to update the\n file before we move on.\n\nStart by reading every file in this Project. Then ask me what I'm\nworking on this week, and we'll plan from there.\n```\n\n\"'Push back' is the line that turns Claude from an assistant into a partner.\"\n> This newsletter grows from your shares. If you've gotten value so far, pass it on to one person who needs it.\n\n\n## Step 5: Build one process file per output type\nFor every repeating deliverable in your business, you need a process file. Not a template. A process.\nA template is a blank doc with headers. A process tells Claude exactly how to think through the work.\nHere are the three I use most. Drop them straight into your Project as separate files.\nProcess file 1: Newsletter issue\n```\n# Newsletter Issue Process\n\nFollow this every time we write a newsletter issue. No skipping steps.\n\nStep 1: Read who-i-am.md and style-rules.md fully before writing\nanything. Confirm you've read them.\n\nStep 2: Ask me three things:\n- What's the topic?\n- What's the one thing I want readers to do differently after reading?\n- Who specifically am I writing this for? (free, paid, or founding?)\n\nStep 3: Suggest 3 angles for the topic. Wait for me to pick one.\nDon't proceed until I choose.\n\nStep 4: Write the title and subtitle first. Wait for approval before\nwriting the body.\n\nStep 5: Write the opening hook. Three sentences max. Wait for approval.\n\nStep 6: Write the body section by section. Pause at the end of each\nsection for a yes/no from me.\n\nStep 7: Run the final check:\n- Does every section pass the style-rules.md test?\n- Is there at least one moment of radical honesty?\n- Is the recap checklist clean and screenshot-worthy?\n- Could this have been written by anyone? If yes, flag it.\n\nStep 8: Deliver the final issue with all visual cues marked.\n```\n![image](https://pbs.twimg.com/media/HMD18e6WAAA9aI9.jpg)\n\nProcess file 2: X post\n```\n# X Post Process\n\nFollow this every time we write an X post.\n\nStep 1: Ask me one question — what's the message in one line?\n\nStep 2: Suggest the best format based on the message:\n- One-liner hook\n- 3-line stacked statement\n- 5-step listicle\n- Thread (only if the message needs more than 280 characters)\n\nStep 3: Write 3 versions. Each one uses a different opening hook style:\n- Bold claim\n- Specific number\n- Personal story line\n\nStep 4: Wait for me to pick one. Then sharpen that one only.\n\nStep 5: Check against style-rules.md. Strip any banned words.\nSend the final version.\n\nNever write more than 3 versions. Never write more than one thread\nper session.\n```\n![image](https://pbs.twimg.com/media/HMD2NssWoAAFTuq.jpg)\n\nProcess file 3: Sponsor reply email\n```\n# Sponsor Reply Email Process\n\nFollow this every time a sponsor email comes in.\n\nStep 1: Read the incoming email I paste. Identify what they want:\n- Pricing inquiry\n- Past work request\n- Booking a specific slot\n- Negotiating a deal\n\nStep 2: Pull my current rate card from what-i-do.md. Pull recent wins\nfrom operating-context.md.\n\nStep 3: Draft a reply that:\n- Answers their direct question in the first sentence\n- Anchors with one specific recent win\n- Names the next clear step (book a call, send rate card, send invoice)\n- Stays under 80 words\n\nStep 4: Flag any email that needs my personal voice — emotional notes,\nweird requests, complaints. Don't draft those. Just summarize and ask\nme to handle it.\n```\n![image](https://pbs.twimg.com/media/HMD2bl4XwAEiktQ.jpg)\n\"The process file is what makes a one-person company feel like a real company.\"\n\n\n## Step 6: Review, don't produce\nThis is the shift most people never make.\nThey use AI to help them produce. The one-person company model uses AI to handle the production and keeps you in the reviewer seat.\nReviewer work looks like this:\n- Is this on-brand?\n- Is this accurate?\n- Is this what I actually want to say?\n- What would make this 10% sharper?\nThat's your only job inside the system. The typing, the structuring, the drafting, the formatting - that's the system's job.\n[VISUAL 3: \"Two Seats\" — Side-by-side split showing Producer seat (stressed, typing) vs Reviewer seat (calm, approving) with the bridging label \"same person, different seat\"]\nTo make this stick, use the review prompt before anything ships.\nThe review prompt\n```\nYou just delivered a draft. Before I read it, run this self-review\nand tell me:\n\n1. Voice check — does this sound like me, or could it be anyone?\n Flag the 3 weakest sentences.\n\n2. Style check — list every place this draft breaks one of my style\n rules from style-rules.md.\n\n3. Honesty check — is there at least one moment where I say what\n didn't work, or call out the overhyped version? If no, flag it.\n\n4. Specificity check — find every vague phrase (\"a lot\", \"many people\",\n \"in general\") and replace with a specific number or example. If you\n can't make it specific, cut it.\n\n5. The \"anyone\" test — could this have been written by anyone? If yes,\n rewrite the weakest section before showing me.\n\nShow me the final version after running all five checks. Then highlight\nwhat you changed and why.\n```\n\n\"The 'anyone test' is the most useful single sentence in this whole system.\"\n\n\n## Step 7: Run a Sunday close\nEvery Sunday night, for 20 minutes, you close the week.\nWithout this, the system slowly drifts. Files get out of date. Claude starts working from the old context. You end up rebuilding the whole thing two months from now.\n![image](https://pbs.twimg.com/media/HMD4-GLXoAAcEbs.jpg)\nThe Sunday close prompt\n```\nWe're closing out the week. Run this with me:\n\n1. Read operating-context.md. Read my weekly brief from Monday.\n\n2. Tell me which of this week's deliverables shipped, which slipped,\n and which we killed.\n\n3. Pull any new decisions I made this week from our conversations.\n Suggest updates to operating-context.md.\n\n4. Pull any new style violations you spotted this week. Suggest 2 new\n rules to add to style-rules.md.\n\n5. Suggest 1 system improvement for next week — one thing we did\n manually this week that should become a process file.\n\n6. Write a fresh draft of next week's weekly brief based on what's\n still open and what's coming next. I'll edit it.\n\nEnd by asking me one question: \"What was the highest-leverage thing\nI did this week, and how can we do more of it?\"\n```\n\n20 minutes. Once a week. That's all it takes to keep the whole system sharp.\n\n\n## A pause before you keep going\nBefore you move on - slow down here.\nYou've just read seven steps. Maybe you're already thinking about which one you'll skip. Don't.\nThe system breaks if you skip the boring steps. The .md files are boring. The Sunday close is boring. The process files feel like extra work. They're not. They're the work.\nMost people will read this issue, open Claude, paste one prompt, and call it a system. That's not a system. That's a vending machine.\nThe people who actually run a one-person company sit down for 90 minutes one Saturday, write the four .md files, set up the Project, and never go back to chaos. That's the whole gap.\nDon't be the vending machine person. Be the 90-minute Saturday person.\n\n\n## Three more advanced prompts to drop in this week\nOnce your Project is set up and your four files are loaded, these three prompts unlock the rest.\nAdvanced prompt 1: The \"build the system, not the one-off\" check\nUse this when you catch yourself doing the same task more than twice.\n```\nI just spent time on [describe the task — example: replying to three\nsponsor inquiries one by one].\n\nBefore I move on, help me decide:\n\n1. Is this a one-time task, or is it something I'll do again in the\n next 30 days?\n\n2. If it's repeating, walk me through what a process file for this\n would look like. Use the same format as my existing process files.\n\n3. Draft a first version of that process file. Include the exact steps,\n the exact files to read, and the exact questions you'd ask me at\n the start.\n\n4. Tell me what I'd save in hours per week if we shipped this process\n file today.\n\nDon't write the full thing yet. Show me a 5-line draft and ask if\nit's worth building out.\n```\n\"Steps 2 and 3 turn a frustration into infrastructure. That's the move.\"\n\nAdvanced prompt 2: The subscriber reply batch\nUse this every Friday. Paste in every reply that came in that week.\n```\nBelow are all the replies I received this week. Treat them as one\nbatch.\n\n[Paste all replies here, separated by ===]\n\nYour job:\n\n1. Categorize each reply into one of these buckets:\n - Quick thank-you (no reply needed)\n - Easy question (draft a short reply)\n - Real feedback (summarize for me, no reply needed)\n - Needs my personal voice (just flag it, don't draft)\n - Sponsor or business opportunity (flag urgently)\n\n2. For the \"easy question\" replies, draft a response in my voice\n (read style-rules.md). Keep each one under 60 words.\n\n3. For the \"real feedback\" replies, give me a 3-bullet summary of\n patterns — what readers are loving, what they're confused by,\n what they want more of.\n\n4. For \"needs my personal voice\" replies, paste the original reply\n and add one line: \"Why this needs you.\"\n\nShow me the batch in a single clean table. I'll approve, edit, and\nsend in one sitting.\n```\n\"The batch turns 45 minutes of inbox grind into 12 minutes of decisions.\"\n\nAdvanced prompt 3: The big decision call\nUse this when you're stuck on a real call. Should I launch this? Take this sponsor? Quit this project?\n```\nI'm stuck on a decision. Here's the situation:\n\n[Describe the decision in 3 to 5 sentences. Include what's at stake,\nwhat your gut is saying, and what's making you hesitate.]\n\nBefore you answer, read who-i-am.md, what-i-do.md, and operating-\ncontext.md. Don't skip this step.\n\nThen, in your reply:\n\n1. Tell me which file from my Project is most relevant to this\n decision and quote the exact line that applies.\n\n2. Lay out the decision as a simple \"if A, then...\" / \"if B, then...\"\n in terms of what changes in my business 90 days from now.\n\n3. Flag any way this decision contradicts what I said I'd do this\n quarter (from what-i-do.md). Don't soften it.\n\n4. Recommend one of the two paths. Give me your reasoning in 3\n sentences max. End with: \"But this is your call, here's what\n you should think about overnight.\"\n\nDon't try to be agreeable. I need the version of you that pushes back.\n```\n\"'Don't try to be agreeable' is the unlock. Without it, you get a yes-man. With it, you get a partner.\"\n\n\n## Best practices and real use cases\nFive ways I actually use this OS every week.\nThe Monday production sprint: Sunday night, I drop my weekly brief into the Project. By Monday morning, I have first drafts of every piece of content for the week. I spend Monday editing only. Tuesday through Friday, everything else gets approved or rejected in under 30 minutes per piece.\nThe subscriber reply batch: Every Friday, replies have piled up. I paste them all into the Project as one batch. Claude categorizes, drafts the easy ones, and flags the ones that need my voice. I approve, customize, and send. 45 minutes turns into 12.\nCompetitive research without the headache: Standing instruction inside the Project: \"Every Wednesday, give me a 5-bullet brief on what's new in the AI tools space that my readers don't know yet.\" I paste in 3 to 5 articles I spotted. Claude synthesizes. I decide what's worth covering. I never read a 31-page report myself.\nThe product page that updates itself: ClaudeKit pages used to be a manual update every time I shipped a feature. Now I have a process file Claude follows. Drop in the feature notes. Out comes an updated, on-brand product description. I read, approve, ship.\nThe big decision call: When I'm stuck on a real call - should I launch this? Should I take this sponsor? - I don't ask my friends. I open the Project and run the Advanced Prompt 3. The Project already knows my goals, my voice, my no-list. The answer is usually clear in five minutes.\nThese are the use cases that matter. Skip the rest until you've nailed these.\n\n\n## The honest part\nI'm going to be honest about something.\nThis system took me four tries to get right.\nThe first time, I wrote the four .md files and forgot about them. Claude never opened them on its own. The output stayed generic. I gave up after two weeks.\nThe second time, I wrote the role prompt but skipped the process files. Every chat was still a one-off. I was just typing fewer words to get the same result.\nThe third time, I had everything except the Sunday close. Three weeks later, the files were so out of date that Claude was giving me advice based on goals I'd already abandoned.\nThe fourth time worked. The fourth time is what you just read.\nDon't expect this to feel magical on day one. Expect it to feel like work for the first week, useful in the second week, and invisible by the third. That's the curve.\n\nHere's the thing nobody tells you about building a one-person company.\nThe hard part isn't the work.\nThe hard part is letting go of the idea that the work has to come from you.\nI spent three months using AI to help me write faster. It helped. But I was still the one producing. The day I handed Claude my style files and said, \"Your job is to protect my voice, not copy it\" - everything changed.\nGood AI use isn't about typing less. It's about staying in the decisions that only you can make. Everything else is infrastructure.\nIf you build this system right, your job stops being \"do the work.\" Your job becomes \"decide what good looks like, and reject anything that isn't it.\"\nThat's the company we're building here.\n\n\n## Recap checklist - in 60 seconds\n→ Build a Claude Project named after your business \n→ Load .md files: who-i-am, what-i-do, style-rules, operating-context\n→ Write a 5-minute weekly brief every Monday \n→ Assign Claude a role with the role prompt - not a task \n→ Write one process file per output type, starting with the one you ship most \n→ Use the review prompt before anything goes live \n→ Run a 20-minute Sunday close every week\nSave this. Screenshot it. Come back to it.\n\n\n## Your gift - as promised\nThe One-Person Company OS - in one Notion workspace.\n→ [Go and get your gift:](https://humzakhalid.substack.com/p/how-to-build-1b-1-person-company?r=31a9dk)\n\nTake 90 minutes this Saturday. Set it up once. Run on it for the next 12 months.\n\nWe tested it in public. Now use it in private.\n- Hamza [��](https://abs.twimg.com/emoji/v2/svg/1f499.svg)��"},"author":{"name":"Hamza Khalid","handle":"humzaakhalid","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/f3eaa0c828977429/avatar-6369bce079a1.jpg","bio":"I create easy-to-read content on (AI in Public) 📨 Collab@aiinpublic.com"},"metadata":{"likes":122,"shares":33,"views":168712,"timestamp":"2026-07-01T09:31:05.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/qfOJoMkPcI/media/0.jpg","type":"image","width":1632,"height":653},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/qfOJoMkPcI/media/1.jpg","type":"image","width":3574,"height":1860},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/qfOJoMkPcI/media/2.jpg","type":"image","width":3576,"height":1860},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/qfOJoMkPcI/media/3.jpg","type":"image","width":3564,"height":1860},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/qfOJoMkPcI/media/4.jpg","type":"image","width":1456,"height":720},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/qfOJoMkPcI/media/5.jpg","type":"image","width":1456,"height":720},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/qfOJoMkPcI/media/6.jpg","type":"image","width":3576,"height":1858}],"aiComments":[]},{"id":"2072677934075424924","shareId":"9H6xX9ylW1","platform":"x","url":"https://x.com/humzaakhalid/status/2072677934075424924","shareMode":"full","createdAt":1783043601666,"postedAt":"2026-07-02T13:45:02.000Z","publishedDate":"2026-07-02T13:45:02.000Z","archivedAt":"2026-07-03T01:53:19.057Z","archivedDate":"2026-07-03T01:53:19.057Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/9H6xX9ylW1","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/MkuHXuDSRi/media/0.jpg","contentType":"post","content":{"text":"this is f*cking gold\n\nSam Altman just predicted a one-billion-person company using AI\n\nAnd I just read a shocking research paper showing how to use a team of AI Agents to build companies without hiring humans.\n\nAnd it's called \"OneManCompany\"\n\nI created a full guide with everything you need.\n\nOnly Claude, a few other tools, and prompts.\n\nClaude helps you build your one billion dollar company exactly like a top executive.\n\nYou don't need to hire humans anymore. You need to build a system where your AI agents will run.\n\nif you want to use highly result-driven skills with one slash activation, try theclaudekit .com\n\nBookmark it before it gets lost in your feed.\n\nRead it now, then check the article below:"},"author":{"name":"Hamza Khalid","handle":"humzaakhalid","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/f3eaa0c828977429/avatar-6369bce079a1.jpg","bio":"I create easy-to-read content on (AI in Public) 📨 Collab@aiinpublic.com"},"metadata":{"likes":49,"shares":9,"views":7743,"timestamp":"2026-07-02T13:45:02.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/MkuHXuDSRi/media/0.jpg","type":"image","width":891,"height":1260}],"quotedPost":{"id":"2072251637742137799","platform":"x","url":"https://x.com/humzaakhalid/status/2072251637742137799","shareMode":"full","content":{"text":"How to Build $1B/ 1 Person Company Using AI\n\nhttp://x.com/i/article/2071895674711658496"},"author":{"name":"Hamza Khalid","handle":"humzaakhalid","avatar":"https://pbs.twimg.com/profile_images/2062191018670501888/M7jUEryG_200x200.jpg"},"metadata":{"likes":122,"shares":33,"timestamp":"2026-07-01T09:31:05.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/MkuHXuDSRi/quoted/0.jpg","type":"image","width":1200,"height":480},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/MkuHXuDSRi/quoted/1.jpg","type":"image","width":1200,"height":625},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/MkuHXuDSRi/quoted/2.jpg","type":"image","width":1200,"height":624},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/MkuHXuDSRi/quoted/3.jpg","type":"image","width":1199,"height":626}]},"aiComments":[]},{"id":"2072616823766368339","shareId":"gFtuve5QRV","platform":"x","url":"https://x.com/DamiDefi/status/2072616823766368339","title":"Anthropic Just Gave Claude Users the Same Research Tools as JPMorgan. Here Is How to Use Them.","shareMode":"full","createdAt":1783043566066,"postedAt":"2026-07-02T09:42:12.000Z","publishedDate":"2026-07-02T09:42:12.000Z","archivedAt":"2026-07-03T01:52:43.391Z","archivedDate":"2026-07-03T01:52:43.391Z","viewCount":6,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/gFtuve5QRV","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/utvIDcvN5t/media/0.jpg","contentType":"article","content":{"text":"Anthropic Just Gave Claude Users the Same Research Tools as JPMorgan. Here Is How to Use Them.\n\nhttp://x.com/i/article/2071290044497686528\n\n---\n\nJamie Dimon Built a Full Treasury Dashboard With Claude in 20 Minutes. Here Is What He Used.\nOn May 5, 2026, Jamie Dimon stood on a stage in New York next to Dario Amodei and described what he had done the weekend before.\nHe had logged into Claude Code on his own. Typed in what he wanted: asset swaps, Treasury bid-ask spreads, investment grade analysis. In twenty minutes, Claude had built him a full dashboard with backup data, research, and analysis accurate enough to use in production.\nThe CEO of JPMorgan. Twenty minutes. On his own.\nThat is the context for what Anthropic launched that same day: ten AI agent templates for financial services, built on Claude Opus 4.7, connected to the data infrastructure that Wall Street firms pay six figures a year to access. FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, LSEG, Moody's, and more.\nMost of these agents are designed for institutional teams. Banks, hedge funds, asset managers deploying Claude at scale across their entire workflow.\nBut three of them are immediately useful for any individual investor who wants to do serious research without a Bloomberg terminal or a financial advisor.\nHere is what they are, how to set them up, and what they actually produce.\n\n## What Anthropic Actually Launched\nThe ten agents are reference architectures, not basic prompts. Each one packages three things: skills (the domain-specific instructions and knowledge for the task), connectors (governed access to financial data sources), and subagents (specialist Claude models called for specific sub-tasks like comparables selection or methodology checks).\nThe agents deploy as plugins in Claude Cowork, as skill bundles in Claude Code, or as cookbooks for Claude Managed Agents running headlessly on Anthropic's infrastructure.\nThe ten agents split into two categories:\nResearch agents (the ones useful for investors):\n- Market Researcher\n- Model Builder\n- Earnings Reviewer\n- Pitch Builder\n- Meeting Preparer\nOperations agents (designed for finance professionals):\n- Valuation Reviewer\n- GL Reconciler\n- Month-End Closer\n- Statement Auditor\n- KYC Screener\nThese updates pair best with Claude Opus 4.7, which leads the Vals AI Finance Agent benchmark at 64.37%, the best of any frontier model on financial tasks.\n\n## Setting Them Up\nStart with the Market Researcher. Not the Model Builder.\nThe Market Researcher requires the least configuration, produces useful output on the first run, and immediately shows whether your setup is working. If the Market Researcher returns a clean portfolio briefing with real analyst data and recent news, the installation is confirmed and the other two agents will work too. If it returns nothing or errors, troubleshoot here before adding the others.\nInstall the Model Builder second, after you have run at least one successful Market Researcher session. The Excel output adds a step that requires either the Microsoft 365 Claude add-in or a manual download, so it is worth confirming the core data access is working first.\nAdd the Earnings Reviewer last. It is the most targeted of the three and works best once you have already used the Market Researcher to identify which positions need thesis checking.\nThere are two ways to install the agents. The easiest is through the Financial Services marketplace inside Cowork. The more configurable path is through Claude Code.\nVia Cowork (easiest):\n> Open Claude Cowork on desktop In the left sidebar, click Plugins Search for Financial Services marketplace Click Add Marketplace Then install the specific agents you want from the marketplace\nVia Claude Code (command line):\n> Add the marketplace first: claude plugin marketplace add anthropics/financial-servicesInstall the core financial analysis skills (required first): claude plugin install financial-analysis@claude-for-financial-servicesThen install the individual agents you want: claude plugin install market-researcher@claude-for-financial-services claude plugin install model-builder@claude-for-financial-services claude plugin install earnings-reviewer@claude-for-financial-services\nYou need a paid Claude plan to use plugins and Cowork. Here is what each tier costs and what it gets you:\n- Pro ($20/month, $17/month annual): includes Cowork, plugins, and the Finance agents. The right starting point for most individual investors.\n- Max 5x ($100/month): same features as Pro, five times the usage headroom. Upgrade here if you hit Pro limits running multi-agent research sessions.\n- Max 20x ($200/month): twenty times Pro's headroom. For daily heavy use.\n- Team ($25/seat/month standard, minimum five seats): for organizations.\nThe agents themselves are free. Plugins and Agent Skills do not cost extra on any paid plan. Some of the premium data connectors (FactSet, S&P Capital IQ, PitchBook) require separate subscriptions from those data providers. Most individual investors will get substantial value from the publicly available data alone: SEC filings, news, earnings transcripts, and free financial APIs.\nOnce installed, the agents appear in Cowork dispatch, skills fire automatically in context, and slash commands become available in your sessions.\nThe slash commands are typed directly in the chat window, not as menu selections. Here is what each one triggers:\n- /comps — runs a comparable company analysis for the ticker you specify next\n- /dcf — builds a discounted cash flow valuation for a named company\n- /earnings — pulls and summarises the most recent earnings call transcript\n- /ic-memo — drafts an investment committee memo for a position you describe\nType /comps NVDA and the agent runs comparables against NVDA's peer group immediately. No additional prompt required. These commands are shortcuts for the most common tasks. For anything more specific than what the command covers, use the natural language prompts in the sections below.\n\n## Agent 1: The Market Researcher\nYour personal analyst covering every stock you own, pulling every relevant data point automatically.\nThe Market Researcher tracks sector and issuer developments, synthesizes news, filings, and broker research, and flags items for review. For an individual investor that translates to: every analyst rating change, every SEC filing, every major news story across your entire portfolio, organized and surfaced without you going looking for it.\nWhat it monitors:\n- News flow, filtered for what actually affects your positions\n- Analyst rating upgrades and downgrades with reasoning\n- SEC filings parsed into plain English (10-K, 10-Q, 8-K)\n- Sector-wide developments that affect your holdings\n- Competitive moves by companies adjacent to what you own\n- Risk signals: regulatory changes, lawsuits, management changes\nWhat \"publicly available data\" actually means without a premium connector subscription:\nWithout a FactSet or S&P Capital IQ subscription, the Market Researcher pulls from SEC EDGAR (all public filings), free financial data APIs (price history, basic fundamentals), web-accessible news, and earnings transcripts from company investor relations pages. This covers most of what an individual investor needs for thesis research.\nWhat it cannot access without a premium subscription: real-time analyst consensus estimates, proprietary broker research reports, and institutional-grade alternative data. If you ask for analyst consensus and Claude cannot access the paid source, it will tell you and work from what is available rather than fabricating numbers.\nPortfolio briefing prompt:\n> I own shares in AAPL, NVDA, MSFT, GOOGL, and AMZN.Give me a full briefing on all five stocks. What happened this week? Any analyst rating changes? Any news I should know about? Flag anything that changes the investment thesis for any of these positions.\nSingle stock deep dive:\n> Give me a complete research report on Tesla (TSLA). Cover: recent earnings, analyst consensus, key risks, competitive position, upcoming catalysts, and whether the current valuation makes sense based on growth.\nSector scan:\n> I am interested in the AI chip sector. Give me an overview of the key players, market share, revenue growth, and which companies are best positioned for the next two to three years. Include NVDA, AMD, INTC, AVGO, and any others worth watching.\nWeekly watchlist ranking:\n> Here is my watchlist: PLTR, SNOW, CRWD, NET, DDOG.Rank them by investment attractiveness right now. For each: current price context, recent developments, analyst sentiment, and the one thing to watch most closely.\nThe most useful setup: give it your full portfolio once and ask for a weekly briefing every Monday morning. The same prompt, run every week, keeps you current without manually checking each position.\n\n## Agent 2: The Model Builder\nFull financial models that investment banks charge clients thousands of dollars to produce.\nGive the Model Builder any publicly traded company and it produces a complete model: revenue and expense projections, DCF valuation, comparable analysis (P/E, EV/EBITDA, P/S), sensitivity tables across different growth and margin scenarios, and a clean assumptions page. In Excel, formatted for review and adjustment.\nThis is the agent that directly replaces the tool individual investors have never been able to access without paying an advisor or subscribing to a professional research service.\nA note on the Excel output before running the first model:\nThe Model Builder produces a downloadable Excel file. How it reaches you depends on your setup.\nIf you have the Claude Microsoft 365 add-in installed (available at claude.ai/integrations or through your organization's Microsoft 365 admin), the model opens directly in Excel with context carrying automatically between Claude and the spreadsheet. This is the cleanest version of the workflow.\nIf you do not have the add-in installed, the model is generated and offered as a file download inside the Cowork window. Click download, open it in Excel or Google Sheets, and review the assumptions tab before interpreting anything else.\nEither way: review the assumptions page first. The model is populated from public data but the growth rates, margin assumptions, and discount rate are defaults that may not reflect your own view of the company.\nFull model build:\n> Build me a complete financial model for Microsoft (MSFT) in Excel. Include five years of revenue and expense projections, a DCF valuation, a comparable analysis versus peers, and a sensitivity table. Show me what the stock is worth at different growth and margin assumptions.\nTwo-stock comparison:\n> I am deciding between CrowdStrike (CRWD) and Palo Alto Networks (PANW). Build me a side-by-side model comparing revenue growth, margins, valuation multiples, and implied upside for both. Which one is cheaper relative to its growth?\nScenario analysis on an existing position:\n> I own Amazon (AMZN). Build me a model that shows what happens to the stock price if: AWS growth slows to 15%, AWS growth stays at 25%, AWS growth accelerates to 35%. Hold everything else constant.\nOne honest point on using this: financial models are only as good as their assumptions. The Model Builder produces the structure and populates it from public data. Review the assumptions it makes before drawing conclusions from the output. Adjust the assumptions that do not match your own research. A model that you have stress-tested is a useful tool. A model you have not looked at is just a formatted spreadsheet.\n\n## Agent 3: The Earnings Reviewer\nEarnings calls are where companies tell you what is actually happening. They are also long, dense with corporate language, and most investors cannot listen to twenty calls a quarter.\nThe Earnings Reviewer reads the full transcript and delivers what matters: what beat and missed, what changed in guidance, whether the thesis you are holding the stock around still holds, and how management's tone has shifted across multiple quarters.\nEarnings recap:\n> Apple (AAPL) just reported earnings. Review the full transcript. What beat and missed expectations? What changed in guidance? Is there anything I should be concerned about?\nThesis check (the most useful version of this prompt):\n> I own Nvidia (NVDA) because I believe AI infrastructure spending will continue to grow for at least three more years. Review their latest earnings call and tell me: does management's commentary support or undermine this thesis? Be specific.\nHere is the kind of output this produces when it works correctly:\nExample Earnings Reviewer output\n> Thesis assessment: Supports with one material complication.Supporting: Jensen Huang described data centre demand as \"insatiable\" and said the transition from general-purpose to accelerated computing is happening \"much faster than we expected.\" Gross margins expanded to 78.4%, indicating no pricing pressure despite accelerating supply. Forward guidance came in above consensus on both revenue and gross margin.Complication: Management flagged that sovereign AI spending, meaning governments building their own national AI infrastructure, is becoming a significant portion of demand. If sovereign spending is pulling forward demand that would otherwise occur over three years, the growth runway you are modelling may front-load rather than extend. This is not a thesis breaker but it is worth tracking quarter over quarter.Net: The three-year AI infrastructure growth thesis is supported by this call. The one question worth adding to your tracking list: what percentage of current demand is sovereign versus commercial, and whether that mix is shifting.\nThat is the difference between reading the earnings call yourself and running it through a thesis-specific prompt. The output is calibrated to your position, not a generic summary.\nMulti-quarter tone analysis:\n> Pull the last four earnings calls for Meta (META). How has management's tone changed over the past year? Are they getting more or less confident about their AI investments? Is ad revenue re-accelerating?\nEarnings season batch:\n> This week, AAPL, AMZN, GOOGL, and META all reported earnings. Give me a summary of each. Which beat expectations? Which missed? Which had the most concerning guidance? Rank them from most bullish to most bearish based on earnings quality.\nThe earnings reviewer works significantly better when you state your thesis first. \"I own this stock because of X. Does this earnings call support or undermine that?\" produces targeted, useful analysis. \"Review this earnings call\" produces a summary you could have read in any financial news outlet.\n\n## Chaining All Three\nThe real value emerges when all three agents run in sequence in a single conversation.\nFull due diligence on a new position:\n> I am considering buying shares of Palantir (PLTR). Run full due diligence:1. Market research: news, analyst ratings, competitive position, key risks 2. Build a full financial model with DCF valuation and sensitivity table 3. Review the most recent earnings call and tell me if the thesis holds 4. Final assessment: buy, hold, or pass, with reasoning\nPre-earnings preparation:\n> Nvidia reports earnings next week. Help me prepare:1. What are consensus estimates for revenue and EPS? 2. What are the three things to watch most closely? 3. Build a quick model showing what the stock is worth at different revenue growth scenarios 4. What did management guide for this quarter on the last call?\nStock screening:\n> Find me five stocks I am probably not looking at that meet these criteria: market cap between $5 billion and $50 billion, revenue growth above 25%, improving margins, strong competitive moat. For each one, give me a quick overview of why it is interesting and what the risks are.\nEach agent builds on what the previous one produced. Research informs the model assumptions. The model gives context for interpreting what management says in the earnings call. The earnings call tells you whether the research thesis is holding.\n\n## What the Other Seven Agents Are For\nThe remaining agents are built for finance professionals and institutional teams:\n- Pitch Builder — creates target lists, comparable analysis, and drafts pitchbooks for investment bankers\n- Meeting Preparer — builds client and counterparty briefing documents before calls\n- Valuation Reviewer — reviews valuations against comparables and methodology for consistency\n- GL Reconciler — reconciles general ledger accounts and performs NAV calculations\n- Month-End Closer — manages the month-end close workflow and task sequencing\n- Statement Auditor — reviews financial statements for internal consistency and audit readiness\n- KYC Screener — handles Know Your Customer document review and compliance preparation\nThese are functional and genuinely useful for their intended audience. For an individual investor, the first three are the ones that matter.\n\n## The Honest Limits\nThese agents produce research and analysis. They do not execute anything. They do not predict the future. They do not replace judgment on whether a thesis is right.\nWhat they do replace is the time cost of pulling data, structuring models, and reading earnings transcripts manually. The analysis that comes back requires the same evaluation you would give any research output: check the assumptions, verify the data sources, apply your own judgment to the conclusions.\nJamie Dimon built his dashboard in twenty minutes. He then presumably applied twenty years of context in interpreting what it showed him. The twenty minutes is what Claude gave him back. The interpretation was still his.\nThese agents give you the same compression. What you do with the output is still the job.\nThe full repository is at github.com/anthropics/financial-services. Everything is open source.\nThis article covers AI research tools, not investment advice. Nothing here is a recommendation to buy or sell any security. Always do your own research and consult a qualified professional before making investment decisions.\nFollow @damidefi on X for daily Claude AI tools, crypto analysis, and the full journey to 100K. Bookmark this. Share it with one person paying a financial advisor for research Claude can now run in under five minutes."},"author":{"name":"Dami-Defi","handle":"DamiDefi","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/0a2105d6ff5a26ad/avatar-1207700a9fde.jpg","bio":"AI. Former marketing lead at global Top 100 company. Crypto since 2018. Sharing strategies that actually work. All posts are NFA | DYOR."},"metadata":{"likes":259,"shares":26,"views":28663,"timestamp":"2026-07-02T09:42:12.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/utvIDcvN5t/media/0.jpg","type":"image","width":1983,"height":793}],"aiComments":[]},{"id":"2072561437407674566","shareId":"QcFeIGJyNx","platform":"x","url":"https://x.com/rubenhassid/status/2072561437407674566","title":"How to learn 80% of Claude Cowork in <20 minutes:","shareMode":"full","createdAt":1783043498464,"postedAt":"2026-07-02T06:02:07.000Z","publishedDate":"2026-07-02T06:02:07.000Z","archivedAt":"2026-07-03T01:51:35.360Z","archivedDate":"2026-07-03T01:51:35.360Z","viewCount":8,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/QcFeIGJyNx","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/0.jpg","contentType":"article","content":{"text":"How to learn 80% of Claude Cowork in <20 minutes:\n\nhttp://x.com/i/article/2071801883531132928\n\n---\n\nI was wrong.\nI wrote “[Cowork](https://ruben.substack.com/p/claude-cowork)” and “[Cowork 2.0](https://ruben.substack.com/p/claude-cowork-20).” My most-shared work ever, across Substack, LinkedIn, and X. A total of 20 million views.\nDo not read them.\nTo be totally fair, I was right about one thing: Claude will become bigger than ChatGPT because of the launch of Cowork in January (I’ve been begging you to switch to Claude since December 21st 2025).\nCowork is the best thing to happen to AI since ChatGPT.\nIf you don’t code, you must be using Claude Cowork.\nSince then, I have written hundreds of guides and two major articles. I shared how to set up Cowork, and that was the latest (outdated) version of it:\n![image](https://pbs.twimg.com/media/HMCFJNOXwAApn7b.jpg)\nThe setup is good…\nUntil you go to enterprises, and they need to collaborate.\nI started giving workshops to enterprises. Teams with hundreds of people using Claude Cowork together. And my files + folder system broke.\nHow Anthropic wants us to use their AI is different from how people actually use it.\nAnd that’s OK.\nBecause I’m not paid by Anthropic.\nSo I can say whatever I want on how to best set up Cowork for you + your team.\nI spent countless hours updating my setup. This guide shows you exactly what & how. By the end of it, you’ll know how to start - and what to absolutely avoid - with Claude Cowork.\nBefore starting, I want you to do two things:\n1. Save this guide & block 20 minutes this week to set up Cowork.\n1. Send it to anyone who still hasn’t tried Cowork (or Claude).\n\n\n# Skip this if you already use Cowork.\nQuick reminder on how to access Claude Cowork:\n1. Go to https://claude.com/download. Download the app on your computer.\n1. You must have a Pro account ($20/month). I pay for the $100/month plan.\n1. Open the app. Click on the Cowork tab at the top between Chat & Code.\n1. You can add Skills and start Projects on the left menu. More about it later.\n1. Make sure to select “Opus” for complex tasks. It’s the smart model.\n1. Once Fable is available, use it. It’s (somehow) even better.\n\n# I. Cowork is many Claude.\nCowork is much better than the normal Claude.\nIn simple terms, why is Cowork better?\nBecause it answers with many Claude at once.\n![image](https://pbs.twimg.com/media/HMCFUvJXkAArMpx.jpg)\n![image](https://pbs.twimg.com/media/HMCFZ3MXIAA_-fV.jpg)\nAfter 8 minutes and many, many Claude sessions, it gave me a great answer. I only asked something once, and Cowork planned everything else.\nClaude Cowork brings many Claude to life to answer your question.\nYou have to imagine separate AIs collaborating to find the solution.\nCowork can perform a task for long minutes, sometimes 30+ min.\nHow to set it up is key, though. And I spent months running Claude Cowork wrong. If you used my old guides, so did you.\nFirst, you must forget about this:\n\n\n# II. Forget about files and folders.\nOK, hold up. I was the files and folders guy.\nI think half of Linkedin is using files and folders on Cowork because of me.\nBut I was wrong.\nFiles and folders suck.\nI told you to set up your Claude Cowork like this, with a folder on your computer and files (about-me files to know about you, and output files for Cowork’s work):\nI would make you create a Cowork folder with files inside.\nAnd I told you to select your folder (with files) when you start a Cowork session.\nBut after months of doing it, I saw two major issues:\n1. Folders are leaking. Even if I told my Cowork to “never check my output folder,” Claude would sometimes do it. And it poisons my context with outdated outputs. Annoying.\n1. Files are impossible to maintain. You have a full-time job. Me too. I can’t spend hours every week updating my “about-me” file. It was nice at first, and then a painful bottleneck.\nInstead, you must focus on two things with Cowork:\n\n\n# III. Skills + Projects. Nothing else.\nSkill = a capability you call. Travels into any chat. Fun fact, it also travels into any AI (like Gemini, ChatGPT…).\nProject = a place you go to. Files + instructions stay loaded, it remembers, and you can share it as one place to collaborate with your team.\nThis is how I work with my team. This is how I work with enterprise teams at GPC. Skills and Projects.\nLet’s go over them one by one and combine them afterward.\n\n## A. What you need to know about Skills.\nSkills are defined by how you use them, right inside any chatbox, with the slash command (for eg, “/linkedin-post”).\nHere’s a skill in action:\n![image](https://pbs.twimg.com/media/HMCtzsyXUAAVJ0f.jpg)\nYou understood how to use a skill.\nBut how do you create a skill? Quite simple.\n1. Either you use the Claude skills /skill-creator.\nYou answer the question Claude will ask you.\nYou test the skill. Save it. And invoke it with / command.\n1. Or you use my free tool I built called “http://makemyskill.com/”.\nI made it to suit how I like to build Claude skills.\nI give it for free because it costs almost nothing to run it.\nOK now in practice it looks like this:\n![image](https://pbs.twimg.com/media/HMCuAs4WwAAYVR5.jpg)\nYou answer the questions.\nCowork goes into the multi-Claude sessions in one. Just wait.\n![image](https://pbs.twimg.com/media/HMCwiznWAAAxwW_.jpg)\nThat was the first way to create a skill.\nBut my favorite way is this one, if you start from scratch:\n![image](https://pbs.twimg.com/media/HMCw1BAWgAA8918.jpg)\nCreate a free account (there is no paid version).\nNow my little tool is creating your Claude skill by scrapping the web much deeper than what Claude usually do.\n![image](https://pbs.twimg.com/media/HMCuh8VW4AI-0u3.jpg)\nGo to Claude Cowork > Customize > Skills.\n![image](https://pbs.twimg.com/media/HMCurK2WMAAPMhr.jpg)\n![image](https://pbs.twimg.com/media/HMCxarLXoAAKD_m.jpg)\n> You can go very broad with skills, giving them tons of data that goes with the prompt. Some of my favorite skills have hundreds of pages of posts, hooks, SOPs, that Claude processes every time for this specific skill.\nI forgot about another way to create a skill. If you’re having a chat with Claude, you can ask it to “make it a skill”.\nStep 1: Go to the top of your chat, on the name of the chat. Click \"Turn into skill\".\nStep 2: Press enter. It will create a skill you can save.\nStep 3: And Claude will create the skill for you. Just wait.\nOK, that was skills. Now projects.\n\n## B. What you need to know about Projects.\nFor Projects, it’s here:\nLeft menu > Projects > New project.\nStart from scratch.\n![image](https://pbs.twimg.com/media/HMCvoaaWMAAuAy6.jpg)\nYou can see your Projects limits at the top right.\n![image](https://pbs.twimg.com/media/HMCvybZXQAACAQI.jpg)\n> Every chat inside a Project has the entire context of this Project. For example, you uploaded all of your clients’ information, the email you both shared, what you agreed on, their favorite past campaign. Claude will remember it on every chat. And it does not poison the other projects (or chats) outside of your Projects.\n\n## C. A skill or a project?\nI have a simple test: \nIt’s a skill you can teach someone → Skill. \nIt’s a specific context (like a client, a campaign) → Project.\nYou must teach your team the difference, otherwise you will end up with hundreds of skills that should be a couple of projects, or dozens of Projects that should be shared Skills.\nI consult companies of 50-500 employees with my team in NYC. Message me if you want a discovery call - that’s how I make money and keep this newsletter free:\n\n## D. Combining the two.\nYes, you can use skills inside projects.\nAnd that’s the best way to use Claude.\n1. Use skills of your most recurring capabilities (eg. make a contract).\n1. Use it inside a project of your client (eg. with his information uploaded).\n1. Connect it before to your Gmail/Outlook and Slack/Teams (to get more context).\nIt looks like this:\n![image](https://pbs.twimg.com/media/HMCv7r0X0AADiF_.jpg)\nHere’s how to share Projects with your team:\n1. Create Projects not with Cowork, but Chat.\n1. Go to the top right, and make sure it’s shared.\n1. Once done, start any task with Cowork. It’s weird, I know.\n1. This Project is inside Chat. You can see your Projects limits at the top right.\n1. Click Share and make it accessible to everyone.\n![image](https://pbs.twimg.com/media/HMC0JT4XMAASRWJ.jpg)\n![image](https://pbs.twimg.com/media/HMC0N2IWYAAGeWb.jpg)\nThat was for Projects.\nNow here’s how to share Skills with your team:\n1. Left menu > Customize.\n1. Skills > Click on one.\n1. Click on “Share” and make sure your workspace is selected.\n![image](https://pbs.twimg.com/media/HMC0YEyWgAA9X5O.jpg)\n\n## E. Not everything is perfect.\nSome problems I hate about this setup:\n1. Skills can auto-fire from their description. I wish Claude could turn this off. I want to activate the skill only when I invoke it.\n1. Projects have a limit on the number of files you can upload. I wish for more (obviously).\nSpeaking of which: Claude has a lot of settings to turn on or off.\nAnd you should turn most of them OFF.\nHere’s how I set up my Claude:\n\n\n# IV. No settings is the best setting.\nI used to build everything around the folder, with files inside.\nIt was literally poison to my Cowork.\nTwo big problems:\n- No one likes to create folders, and maintain them.\n- No one likes to create files, and maintain it.\nI also believe something, and everyone says the opposite: you shouldn’t give too many instructions to an AI.\nOnce you give too much context to Claude, it falls into the same answers every single time. No more creativity.\nIt’s OK when you need to draft the same contract.\nIt’s not OK to brainstorm and find creative solutions.\nIf you give too much context to Claude, it’s going into circles.\nSo here’s the very simple setup I have today:\n#1. No more folders. And I use Opus 4.8 (High) most of the time. Max effort when in need, or Fable-5 once it’s available.\n#2: In my General settings, the minimum possible. Don’t give any info.\n![image](https://pbs.twimg.com/media/HMC02T1WMAAe_2R.jpg)\nYou can find Connectors in the + to connect apps like Gmail.\n![image](https://pbs.twimg.com/media/HMC07quXwAAqmjZ.jpg)\n![image](https://pbs.twimg.com/media/HMC2BClW0AE35ge.jpg)\n![image](https://pbs.twimg.com/media/HMC-3y2W8AAgpgF.jpg)\n#6: Keep your Global Instructions free of context.\nYou have now completed a setup of basically… nothing. Everything is mostly turned off. So what should you do instead?\n1. Skills for recurring tasks.\n1. Projects for recurring projects.\n1. AskUserQuestion to prompt better.\nYou don’t know the third one.\nBut it’s my favorite Claude trick:\n\n\n# V. This is still my #1 trick.\nYou don’t know how to prompt Claude.\nOr any LLMs to be honest.\nIt takes time, practice, experience, understanding of how the technology works. It’s not meant to say the truth, for example, but answering (which leads to hallucination, but it can be solved).\nThis simple trick will fix most of your prompt problems:\n> \"I need to [do a task, for eg, \"write a PR brief for the launch of my company] for [success criteria, for eg, \"to get more leads\"]. First, use the tool AskUserQestion to get more info.\"\nClaude creates a form you just have to answer.\nOnce you answer, Claude will think more about how to reply.\n![image](https://pbs.twimg.com/media/HMC2aqnXEAA0gjn.jpg)\nThe AskUserQuestion trick is infinite.\nIt’s not just the first prompt. You can follow up with it forever.\nAnd Claude will keep interviewing you until you’ve given so much context it’s impossible to fail.\n\n\n# VI. You really don’t need folders.\nI know you are pissed at me because you’ve heard me saying files and folders are so cool.\nAnd I really believed it.\nBut then I checked how I use Claude, and I just don’t open Claude’s files.\nLet me give you an example.\nWhen I create a spreadsheet, an Excel, using Cowork, I always export it with one button to my Google Drive (with Google Sheet).\nSame goes to every document (text, PDF, slides…).\nHere’s how it works:\nStep 1 - I like using this template to make a clean Excel.\n> \"Create an Excel spreadsheet from:\n[DATA: file path, folder, or pasted data, best is to upload data].\n\n### Purpose:\n[Who uses this and what decision/task it supports — 1 sentence.]\n\n### Sheets needed:\n- “[Sheet name]”: [columns, what each row represents, any calculations or formulas]\n- “[Sheet name]”: [e.g., summary with totals, pivot, charts — or remove if only one sheet]\n\n### Formatting:\n[Currency/date formats, conditional highlighting, frozen header row, totals row — pick what applies.]\n\nBefore building, list your top 10 assumptions so I can sanity-check them, then execute.\"\nHere’s an example:\n> \"Create an Excel spreadsheet: Customer support capacity and cost model for a fast-growing B2B SaaS company, covering Q3 2026 – Q2 2027.\n\nPurpose:\nLeadership-ready model to decide when to hire support agents, where automation helps most, and how support costs scale as customer count grows.\nBusiness context:\n\n* Company sells project-management software to SMB and mid-market customers.\n* Current customer base: 4,800 accounts, growing 8% month-over-month.\n* Support channels: email, live chat, phone, and self-serve help center.\n* Ticket volume depends on customer count, onboarding activity, product complexity, and seasonality.\n* Make your own assumptions for tickets per account, channel mix, first-response time targets, handle time, agent productivity, automation deflection, salary costs, software costs, escalation rates, and customer satisfaction impact.\n* Every assumption lives as a labeled, editable input on the Assumptions sheet — never hardcoded inside a downstream formula — so the COO can pressure-test staffing and automation decisions live.\n\nSheets needed:\n* “Assumptions”: all editable inputs grouped by Customer Growth, Ticket Demand, Channel Mix, Agent Productivity, Automation, Cost Structure, and SLA Targets\n* “Ticket Forecast”: monthly ticket volume by channel, including new-customer onboarding tickets, recurring support tickets, escalations, and deflected tickets\n* “Capacity Plan”: required agent hours, available agent capacity, hiring needs, utilization %, backlog risk, and SLA coverage by month\n* “Cost Model”: support labor cost, software/tooling cost, outsourcing cost, automation investment, cost per ticket, and cost per customer\n* “SLA & CX”: first-response time, resolution time, backlog, CSAT estimate, escalation rate, and churn-risk indicator\n* “Dashboard”: single-screen executive summary with KPI tiles for total tickets, required headcount, support cost/customer, SLA attainment, automation savings, and CSAT trend; include charts for ticket volume by channel, capacity gap, and support cost over time\n* “Scenarios”: Base / Growth Surge / Automation-Heavy toggle that flexes customer growth, ticket volume per account, automation deflection, and agent productivity\n\nFormatting:\nUSD currency with no decimals, percentages where relevant, frozen header rows, clean executive-friendly palette, conditional formatting for SLA misses and utilization above 85%, monthly columns grouped by quarter, totals row on every model sheet, and named ranges for major assumptions so formulas are readable.\nBefore building, list your top 10 assumptions so I can sanity-check them, then execute.\"\nI just received a multi-tab Excel.\nAnd the magic button is the Google Drive integration:\n![image](https://pbs.twimg.com/media/HMC3LzAXgAAS9Jj.jpg)\nAnd if you need to edit it, you have two choices from here:\n1. Go back to Cowork and ask what you want. And if you don’t know what to ask, ask Claude to ask you. Using the AskUserQuestion tool I told you about.\n1. Or use the (free) ChatGPT extension inside Google Sheets. Here's how:\nGo to the top → Extensions → ChatGPT (you might have to add it first).\nIt will open a sidebar. I like using the Heavy mode.\nI tested every single ways to create spreadsheets with AI. Cowork is the best.\nAnd it’s not even the only cool part of Cowork.\nYou can onboard anyone of your company much more easily with Claude:\n\n\n# VII. Onboard with Claude.\nGo to your Claude app. Create a new project.\n1. Add all of the necessary resources, usually the company’s mission, deck, some interviews of the CEOs or C-levels…\n1. If you can add best practices from support, they usually cover everything about the business a new employee should know.\n1. Invite them to the new Claude Project.\n1. Create a skill on top of it that is called /answer-onboarding with how to use the data inside the Claude Project to answer anyone doing an onboarding.\n1. Now as soon as a new recrue needs an answer, they start by going to the Claude Project, and do /answer-onboarding + ask the question.\n1. Claude will give a great answer.\nI have a rule in my company:\n- Please ask me as many questions as possible. Failing is winning.\n- Never ask me a question before giving me a potential answer.\n- Find the potential answer by doing what I just said: project + skills invoking how to (potentially) answer the new person.\nIt forces people to go from “I’m new, I can’t think for myself” to “I work actively to find the answer, but I still need help”.\nIt’s very different.\nAfter this you’ll want to run everything through Cowork.\nDon’t.\n\n\n# VIII. When Cowork is the wrong tool.\nYou should use…\n- Chat: nothing to build, nowhere to keep it, a question, a quick rewrite.\n- Claude Code: the deliverable is software, code, something to build.\n- Claude Cowork: everything document-shaped (decks, analysis, client work).\n![image](https://pbs.twimg.com/media/HMC4U34XIAECYA6.jpg)\nClaude is extremely powerful.\nSo powerful that its biggest blocker is most of the time… you.\n\n\n# IX. You are slowing us down.\nThis is a typical Cowork session:\n1. You type a prompt. 30 seconds.\n1. Cowork reads your skills. 50 seconds. It generates a plan. 50 seconds. It asks you clarifying questions using AskUserQuestion. 50 seconds.\n1. Now you answer those questions. You click some options. Fine.\nBut sometimes you need to type a custom answer (and custom answers are where the best outputs come from).\nSo you stop. Think. Type.\n60 seconds per answer. Maybe 2 minutes.\nAcross 8 questions, that’s 8-15 minutes of you being the slow part.\nBut Cowork can read 100,000 words in 15 seconds. It can build a spreadsheet in 90 seconds. And it has to wait for you to type at 60 words per minute.\nBut you could be faster. Because you speak at 150 words per minute.\nSide note [⚠️](https://abs.twimg.com/emoji/v2/svg/26a0.svg)\nI know it sounds ridiculous to optimize everything for speed.\nBut talking instead of typing has another massive benefit: it sounds natural. It’s you, your voice.\nAnd your brain thinks very differently when it has to talk.\nDid you realize how good your ideas are, your flow is, when you’re talking to a colleague about solving a problem?\nWe want the same flow state here.\n\n## How to set up Wispr flow.\nQuick reminder for those who missed it: Wispr Flow is a dictation tool.\nYou hold a key, talk, release. Your words appear wherever your cursor is.\nAnywhere on your computer. Including inside Cowork’s chat box.\nWhat makes Wispr Flow different: its accuracy.\nNear perfection, every single time.\nWispr Flow + Claude Cowork is the perfect match to counter a common problem when using AI: you steer the conversation, and make sure you are in flow state.\nInstead of typing “I need a Linkedin post,” you start talking “I recently found out about… and I want to share more about… but first I need to make sure that… so maybe we should start covering… and end up with… as a conclusion”.\n\n## 1. The initial prompt: I speak it.\nI say this out loud. Wispr types it.\nThe point isn’t to be faster, but we end up giving much more context when we talk (we are yappers by nature), rather than with a lazy typed prompt.\nAnd the more context, the better.\n\n## 2. AskUserQuestion answers: I speak those too.\nCowork generates a form. Most options I just click.\nBut when I need to add context (”make it more direct, she’s a CEO who hates fluff, and reference the ROI data from the last call”), I dictate that instead of typing it.\nI don’t self-edit while speaking. I dump my thinking.\nCowork figures out what matters.\n\n## 3. Feedback and pivots: spoken.\nWhen Cowork produces something that’s off, I used to type feedback like: “Tone is wrong. Make it less formal.”\nNow I say: “The tone is too stiff. I want it to sound like I’m texting a friend who happens to run a 200-person company. Keep the data but make it casual. Only redo section 2.”\nSpoken feedback is richer because I am a yapper.\nHow to download Wispr Flow for free.\n1. Go to https://wispr.ai/. Download Wispr Flow. Install it.\n1. The free tier is capped at 2,000 words/week - perfect to know if you like it.\n1. Choose your favorite keystroke to activate Wispr Flow. I personally love “Shift”, the little arrow below “Enter”.\n1. Go to any app, hold your keystroke (like shift) and… talk.\n1. You now type 4x faster than before. Congratulations.\n1. You get a [free month](https://ruben.substack.com/p/upgrade) when you pay for my newsletter (you don’t have to).\nThere’s no “integration” to configure. Wispr types wherever you type.\nHere’s how it looks:\n\nMy Cowork literally spawned 5 running agents. Mini-Claudes.\nNow there might be a problem.\nYoull end up using Cowork so much you will spend a lot of money.\nSo here is how to save money in Cowork:\n\n\n# X - How to save money in Cowork.\nThe $20 paid plan of Claude will give you credits (called tokens).\nBut you will use them up very fast.\nAt the scale of a company, saving tokens = saving thousands of dollars!\nThis section is about saving as many credits (= tokens) as possible.\n\n## 1 - Restart your conversation. Don’t send a follow-up.\nFirst, Claude doesn’t count messages. It counts tokens. Every time you send a message, Claude re-reads the entire conversation history.\nMessage 30 costs 31x more tokens than message 1.\nWhen Cowork gets something wrong, you want to type “No, I meant...” and send another message. Don’t.\nEvery follow-up stacks on top of the full conversation history. Claude re-reads all of it every turn. At ~500 tokens per exchange, 20 messages burns 105K tokens. 30 messages burns 232K.\nInstead: click “Restart the conversation from here” on a previous message (much higher ideally so you do save tokens).\nIt's at the bottom right of everyone of your prompt.\nIn Cowork, you can also edit a previous message.\nClick on Edit on your own prompt. It will delete whatever comes next.\nAnd you can edit your prompt to start over (from there).\nSo when something goes wrong, edit your prompt with a better prompt or restart the conversation higher.\nIt’s the biggest hack!\n\n## 2 - Start a fresh session every 20 messages.\nLong conversations are token furnaces.\nOne developer tracked his usage and found 98.5% of tokens were spent re-reading history. Only 1.5% went to the actual output.\nWhen a Cowork session gets long: ask Claude to summarize everything, copy it, start a new session, paste the summary as your first message.\nThere is no specific prompt, but I like this one:\n> \"This Cowork session is getting too long. I need you to create the best summary possible as an .md file so I can start a new Claude Cowork session without losing too much context.\"\nYou keep the context. You lose the bloat.\n\n## 3 - Batch your tasks into one message.\nThree separate prompts = three full context reloads.\nOne prompt with three tasks = one reload.\nInstead of: “Summarize this article” then “List the main points” then “Suggest a headline” write: “Summarize this article, list the main points, and suggest a headline.”\nUsing Wispr Flow helps.\n\n## 4 - Use Sonnet (not Opus) for quick tasks.\nGrammar checks, brainstorming, formatting, short answers. Sonet handles all of this at a fraction of the cost. And technically, Haiku is even cheaper.\nSave Opus + Extended thinking for the work that actually needs it. Drafts and simple tasks on Sonnet-Haiku free up 30-70% of your budget for deep work.\nWhat about “Mythos” or “Fable-5” level of Claude?\nFable-5, called a “Mythos-level,” is even better than Opus.\nIt is obviously for your hardest tasks possible.\nI’ll be honest, I was spamming Fable-5 requests whenever I could.\nBut you know it’s the most token-intensive model.\nSo there is no right or wrong answer. I am paying $100 a month for my Claude, and I have never run into any problems. We don’t use it for coding, so we are using fewer tokens than typical people.\n\n## 5 - Delete your ABOUT ME file.\nI wrote extensively on building your “about-me” file.\nKeep it. But turn it into a skill:\n![image](https://pbs.twimg.com/media/HMC8WANXYAA0O5t.jpg)\nAnd I can now invoke it whenever I want.\n\n## 6 - Spread your work across the day.\nThis one is impossible to actually do in practice, but it works.\nClaude uses a rolling 5-hour window. If you burn your entire limit in one morning session, most of your daily capacity goes unused.\nSplit into 2-3 sessions: morning, afternoon, evening.\nBy the time you come back, your previous usage has rolled off.\nAnd avoid peak hours (5-11 AM Pacific on weekdays) when the same query costs more against your limit.\nCool in theory, but I never do it ahah.\nIf I want to work, I work.\nThat’s why I pay for the $100/month paid plan.\nSo far, no problem (but I had a lot on the $20 plan).\nI think a couple of dollars per day is worth it to have a godlike intelligence by my side. Your call.\n\n\n# XI. Your first 20 minutes.\n- 0–3: empty your settings, turn everything off.\n- 3–8: build your first Skill (something you keep doing).\n- 8–13: create your first Project (upload PDFs, excels… from a client).\n- 13–17: use your skill inside your project to see how it feels.\n- 17–20: download my Claude Skills library by subscribing for free.\nI give the library for free in my welcome email.\nIt’s all for free for new subscribers.\nCreating skills should be your priority once you set up Cowork.\nIf you want (yet) another way to create skills, here’s a cool last example:\nI saw this [tweet](https://x.com/jakecastilloooo/status/2070168892098416951?s=20) praising a lot of experts.\nGo to Chat (not Cowork) and turn on “Research”. I then copied and pasted the tweet.\nIt generated a long research on the topic after a deep web search.\nClaude decided to create 7 skills out of it.\nSky is the limit. Create many skills, test them, improve them.\n\n\n# PS: You don’t want your company to be left behind.\nReading about AI is useful.\nBut it won’t change how your company works.\nIf you have 100+ people on your team, adopting AI is not a “send them a newsletter” problem. It is a systems problem.\nYour people have different roles, clients, tools, workflows, permissions, data, risks, and habits. My newsletter, sent to 750,000 people, can’t do that.\nThat is how I help companies.\nI run tailored AI workshops for enterprise teams. I am not alone, we are 20 people in New York City to support our clients.\nWe map where AI can actually save time, where it should not be used, and how your people can adopt it without creating chaos.\nIf your company has 100+ people, message me.\nMy team & I will take care of you in priority.\n\n\n# A message from the author, Ruben.\nThis article exists because 89,000+ people decided AI is too important to leave aside. Not only that, but they shared it around them. They understand they are the sum of the 5 people around them. So better have them using AI.\nIf this helped you — or if it’ll help someone you know — forward it to them. That’s how this grew. Just readers like you sending it to people like them.\nAnd if you're new here, follow me on X →[@rubenhassid](https://x.com/@rubenhassid) (also free!)"},"author":{"name":"Ruben Hassid","handle":"rubenhassid","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/f3a4e34aacb052ff/avatar-8e73bcbc3293.jpg","bio":"some people read my https://t.co/uN7GJvEWtp"},"metadata":{"likes":153,"shares":20,"views":11991,"timestamp":"2026-07-02T06:02:07.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/0.jpg","type":"image","width":4096,"height":1638},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/1.jpg","type":"image","width":1456,"height":910},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/2.jpg","type":"image","width":1456,"height":913},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/3.jpg","type":"image","width":1456,"height":913},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/4.jpg","type":"image","width":1456,"height":910},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/5.jpg","type":"image","width":1456,"height":739},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/6.jpg","type":"image","width":1456,"height":913},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/7.jpg","type":"image","width":1456,"height":1820},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/8.jpg","type":"image","width":1456,"height":1031},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/9.jpg","type":"image","width":1456,"height":503},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/10.jpg","type":"image","width":1456,"height":730},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/11.jpg","type":"image","width":1456,"height":913},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/12.jpg","type":"image","width":1456,"height":950},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/13.jpg","type":"image","width":1456,"height":756},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/14.jpg","type":"image","width":1456,"height":913},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/15.jpg","type":"image","width":1456,"height":913},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/16.jpg","type":"image","width":1456,"height":806},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/17.jpg","type":"image","width":1456,"height":514},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/18.jpg","type":"image","width":1456,"height":912},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/kjA4WQpKD1/media/19.jpg","type":"image","width":1456,"height":913}],"aiComments":[]},{"id":"2072670403747537159","shareId":"Uy8rBXapOM","platform":"x","url":"https://x.com/wei_wang/status/2072670403747537159","shareMode":"full","createdAt":1783043452885,"postedAt":"2026-07-02T13:15:07.000Z","publishedDate":"2026-07-02T13:15:07.000Z","archivedAt":"2026-07-03T01:50:51.299Z","archivedDate":"2026-07-03T01:50:51.299Z","viewCount":2,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/Uy8rBXapOM","contentType":"article","content":{"text":"有人来问我平常都在用哪些 Claude skills，分享几个我最近常用的：\n\n1. caveman\n\n让 Claude 回复只留重点、废话全砍。最大好处是超省 token，长对话成本直接砍一截。\n\n适合那种你已经知道自己要干嘛，只需要它快速执行，不想听它写小作文的场景。\n\n2. graphify\n\n把整包项目（代码、SQL、文件）变成可查询的知识图谱。问它“这段逻辑在哪、谁调用谁”秒答。\n\n项目一大之后，人脑根本记不住所有依赖关系，这种东西就是给 AI 装一个项目地图。\n\n3. codeburn\n\n一个仪表盘，让你看 AI 把 token 和花费烧在哪，不再黑箱。\n\n用 agent 最怕的就是钱不知道花哪了，它可以让你看清楚到底是哪一步在疯狂烧 token。\n\n4. claude video\n\n让 Claude 真的看懂视频，自动抓影格加转语音，再回答视频里发生什么。\n\n适合分析课程、会议、产品 demo，不用自己一帧一帧看，直接把视频变成可提问资料。\n\n5. open design\n\n本地开源版的设计工作室。用你自己的设计系统，让 AI 直接生网页、简报、图片原型，风格还一致。\n\n最有用的是能避免 AI 每次都乱换审美，让产出的东西看起来像同一个品牌做的。\n\n6. browser-harness\n\n让 AI 更稳地用 Playwright 自动操作网页，挡掉一堆自动化会踩的雷。\n\n只要你让 agent 操作网页、测试流程、跑表单，这类东西基本就是刚需，不然很容易一路点崩。\n\n7. gsap\n\n教 AI 正确写 GSAP 网页动画，不会乱兜一通。\n\n做网页动效的时候非常有用，不然 AI 很容易写出一堆能跑但很丑、很卡、很难维护的动画。"},"author":{"name":"Wei","handle":"wei_wang","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/758767c048c1ff39/avatar-50b28532a6bc.jpg"},"metadata":{"likes":38,"shares":9,"views":2727,"timestamp":"2026-07-02T13:15:07.000Z"},"media":[],"aiComments":[]},{"id":"2072774041535185119","shareId":"r8iioMSpNi","platform":"x","url":"https://x.com/angeldot_/status/2072774041535185119","shareMode":"full","createdAt":1783042537238,"postedAt":"2026-07-02T20:06:56.000Z","publishedDate":"2026-07-02T20:06:56.000Z","archivedAt":"2026-07-03T01:35:34.520Z","archivedDate":"2026-07-03T01:35:34.520Z","viewCount":9,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/r8iioMSpNi","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2065483060523491330/img/xBD2J00WmY35z9sg.jpg","contentType":"post","content":{"text":"EL CREADOR DE CLAUDE CODE YA NO ESCRIBE PROMPTS EN FABLE 5\n\nBoris Cherny, la persona que construyó Claude Code, dijo esto en una entrevista:\n\n\"Con Fable 5 no escribo prompts. Escribo loops, y los loops hacen el trabajo. Mi trabajo es escribir loops.\"\n\nMientras la mayoría de devs sigue puliendo el prompt perfecto, él ya pasó a otra cosa.\n\nEn 30 minutos enseña su setup diario real:\n\n→ Claude Code (Fable 5) + loops + workflows dinámicos\n\n→ Su flujo de trabajo de cada día\n\n→ Cómo aplicarlo desde hoy\n\nVale más que cualquier curso de 500$.\n\nGratis. Subtitulado a español e inglés.\n\nY después lee este artículo:\n\nTodo lo que necesitas saber sobre loops para aplicar lo que cuenta Boris ↓\n\nGuárdate esto. 🔖\n\nYa tienes plan este finde. https://x.com/nicos_ai/status/2065485711491334384/video/1"},"author":{"name":"angel","handle":"angeldot_","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/00f0a302a7c6e29b/avatar-0d2fe082a9ec.jpg","bio":"CTO Accesia +300K ARR | AI & tech | growth @aiscwork | @eComrads"},"metadata":{"likes":99,"shares":14,"views":12283,"timestamp":"2026-07-02T20:06:56.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2065483060523491330/vid/avc1/1280x720/cImywYxSmS7KuqQa.mp4?tag=27","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2065483060523491330/img/xBD2J00WmY35z9sg.jpg","type":"video","width":2560,"height":1440,"duration":1792}],"quotedPost":{"id":"2065145486890164524","platform":"x","url":"https://x.com/angeldot_/status/2065145486890164524","shareMode":"full","content":{"text":"Cómo crear Loops en Fable 5\n\nhttp://x.com/i/article/2064911650314911746"},"author":{"name":"angel","handle":"angeldot_","avatar":"https://pbs.twimg.com/profile_images/2042194820643291136/a7VPl8re_200x200.jpg"},"metadata":{"likes":419,"shares":78,"timestamp":"2026-06-11T18:53:47.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/q6Yr6ooVjN/quoted/0.jpg","type":"image","width":1200,"height":480},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/q6Yr6ooVjN/quoted/1.jpg","type":"image","width":1200,"height":695},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/q6Yr6ooVjN/quoted/2.jpg","type":"image","width":1200,"height":565},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/q6Yr6ooVjN/quoted/3.jpg","type":"image","width":1200,"height":547}]},"aiComments":[]},{"id":"2072749611471835229","shareId":"xTvgM3cBeL","platform":"x","url":"https://x.com/sairahul1/status/2072749611471835229","shareMode":"full","createdAt":1783042527141,"postedAt":"2026-07-02T18:29:51.000Z","publishedDate":"2026-07-02T18:29:51.000Z","archivedAt":"2026-07-03T01:35:24.097Z","archivedDate":"2026-07-03T01:35:24.097Z","viewCount":1,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/xTvgM3cBeL","thumbnail":"https://pbs.twimg.com/media/HMPPT0saQAAVC1e.jpg","contentType":"post","content":{"text":"This is the best site on the internet to find agent loops you can use right now.\n\nFree. Completely.\n\nMost AI engineers are still writing these from scratch.\n\nhttps://signals.forwardfuture.com/loop-library/\n\nBookmark this site.\n\nThen read this ↓ https://t.co/H5vkrn3UJ2"},"author":{"name":"Rahul","handle":"sairahul1","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/1f8e3589cf94d186/avatar-aae727dbc242.jpg","bio":"Building with AI. Sharing what's wild, what's practical, and what's next. Founder of https://t.co/IfYUA3MgDK and https://t.co/jeqp1MyNtk"},"metadata":{"likes":111,"shares":13,"views":16501,"timestamp":"2026-07-02T18:29:51.000Z"},"media":[{"url":"https://pbs.twimg.com/media/HMPPT0saQAAVC1e.jpg","type":"image","width":1692,"height":1588}],"quotedPost":{"id":"2072258045460226373","platform":"x","url":"https://x.com/sairahul1/status/2072258045460226373","shareMode":"full","content":{"text":"20 Loop Design Patterns Every AI Engineer Should Know\n\nhttp://x.com/i/article/2072234929048649728"},"author":{"name":"Rahul","handle":"sairahul1","avatar":"https://pbs.twimg.com/profile_images/1999441729800536065/6-3Y5xyR_200x200.jpg"},"metadata":{"likes":265,"shares":55,"timestamp":"2026-07-01T09:56:33.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/D7y2lBjV2C/quoted/0.jpg","type":"image","width":1200,"height":480},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/D7y2lBjV2C/quoted/1.jpg","type":"image","width":1199,"height":675},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/D7y2lBjV2C/quoted/2.jpg","type":"image","width":1199,"height":675},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/D7y2lBjV2C/quoted/3.jpg","type":"image","width":1199,"height":675}]},"aiComments":[]},{"id":"2072579279926481350","shareId":"pFDTxqItmb","platform":"x","url":"https://x.com/cyrilXBT/status/2072579279926481350","shareMode":"full","createdAt":1783042426775,"postedAt":"2026-07-02T07:13:01.000Z","publishedDate":"2026-07-02T07:13:01.000Z","archivedAt":"2026-07-03T01:33:41.560Z","archivedDate":"2026-07-03T01:33:41.560Z","viewCount":3,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/pFDTxqItmb","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/ezJH72zew2/media/0.jpg","contentType":"post","content":{"text":"GITHUB JUST CREATED AN OFFICIAL CERTIFICATION FOR THE MOST IN-DEMAND DEVELOPER ROLE OF 2026.\n\nIt is called Agentic AI Developer.\n\nGH-600.\n\nAnd it is the first formal signal that running AI agent teams is now a recognized engineering discipline with a credential behind it.\n\nNot a prompt engineer.\n\nNot a vibe coder.\n\nAn Agentic AI Developer.\n\nThe person who operates, supervises, and integrates AI agents across the entire software development lifecycle.\n\nThe person who knows where agents fail in production.\n\nThe person who understands how to build autonomous workflows that do not introduce catastrophic failure modes into CI/CD pipelines.\n\nThe person every engineering team is going to need and almost none of them have right now.\n\nGitHub certifying this role changes the hiring conversation permanently.\n\nBefore GH-600: \"Do you work with AI agents?\" is an interview question with no standard answer.\n\nAfter GH-600: the credential tells the hiring manager exactly what you know and what you can do before the interview starts.\n\nThe engineers who get certified in the first wave of GH-600 will have a credential for a role that has more demand than supply for the next 3 to 5 years.\n\nThe engineers who wait until it is mainstream will be competing with everyone who moved first.\n\nIf you are already working with GitHub Copilot or building agent-driven workflows you are already doing this job.\n\nGH-600 is how you prove it.\n\nBookmark this."},"author":{"name":"CyrilXBT","handle":"cyrilXBT","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/be78214b26751fa6/avatar-d0634120b3b2.jpg","bio":"AI • Tech • Crypto | Sharing what I see before everyone else does."},"metadata":{"likes":476,"shares":65,"views":40167,"timestamp":"2026-07-02T07:13:01.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ezJH72zew2/media/0.jpg","type":"image","width":1200,"height":612}],"quotedPost":{"id":"2072139907615560054","platform":"x","url":"https://x.com/cyrilXBT/status/2072139907615560054","shareMode":"full","content":{"text":"Claude cowork just made the one person company actually possible in 2026 - Full Course \n\nhttp://x.com/i/article/2071072979262312448"},"author":{"name":"CyrilXBT","handle":"cyrilXBT","avatar":"https://pbs.twimg.com/profile_images/2035229727414534145/aWap3Jbq_200x200.jpg"},"metadata":{"likes":197,"shares":27,"timestamp":"2026-07-01T02:07:07.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/ezJH72zew2/quoted/0.jpg","type":"image","width":1200,"height":480}]},"aiComments":[]},{"id":"2072421728396226902","shareId":"ZtzfLJdVcu","platform":"x","url":"https://x.com/gkisokay/status/2072421728396226902","shareMode":"full","createdAt":1783002926834,"postedAt":"2026-07-01T20:46:58.000Z","publishedDate":"2026-07-01T20:46:58.000Z","archivedAt":"2026-07-02T14:35:21.808Z","archivedDate":"2026-07-02T14:35:21.808Z","viewCount":1,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/ZtzfLJdVcu","contentType":"article","content":{"text":"Fable is back, but unless you use a chain of prompts, you'll never maximize its effectiveness.\n\nBookmark this for later, I still refer to it for all of my builds.\n\nHere's the 1st of 5 prompts that enables Fable to audit any repo:\n\n\"You are auditing this repository as a principal-level engineer. Produce an honest audit and a prioritized, actionable improvement plan. This is analysis only: do not modify any files — the deliverable is the audit document itself.\n\n\n\nThe audit will be used to decide what to fix, in what order, before further development. Recommendations must fit the project's actual maturity and conventions — advice calibrated to a production service is wrong for a prototype, and vice versa. An honest \"this area is healthy\" is more useful than padded findings.\n\n\n\n\n\nWork in four phases, in order. Ground every claim in actual files — cite file paths and line numbers. If you can't verify something, say so explicitly rather than guessing. Distinguish facts (\"no error handling: src/api/client.ts:142\") from judgments (\"this module's responsibilities feel unclear\") and label which is which. Prefer 15 high-confidence findings over 50 speculative ones. If the repo is large, go deep on the core ~20% of code that does most of the work and note which areas got lighter review. Use parallel subagents or parallel file reads where it speeds up discovery.\n\nPhase 1 — Discovery (read before judging). Map the directory structure; identify project type, languages, frameworks, entry points, and main data/control flow. Read manifests, lockfiles, build/CI config, and docs. Determine the project's purpose, intended users, and maturity. Note existing conventions (naming, boundaries, error handling, test style) so recommendations fit the culture rather than fighting it.\n\nPhase 2 — Audit. For each dimension below, record per finding: what, where (file:line), why it matters (concrete consequence), and severity (Critical/High/Medium/Low). Also list genuine strengths — they decide what to preserve. If a dimension is healthy, say so in one sentence and move on.\n\nPhase 3 — Strategy. Identify the 3–5 themes that explain most findings. For each: target state and the principle behind it. State explicit trade-offs, including what NOT to fix and why. Define measurable \"done\" signals (e.g., \"CI fails on lint errors\", \"core module coverage ≥ 80%\").\n\nPhase 4 — Task plan. Discrete tasks, each with: title, one-paragraph description, files affected, acceptance criteria, effort (S <2h / M half-day / L 1–2 days / XL needs breakdown), risk of the change itself, and dependencies. Order into milestones: M0 safety net (tests around critical paths, CI gates) → M1 critical security/correctness fixes → M2 high-leverage improvements → M3 quality and polish. Flag quick wins (high impact, S effort) separately. For the top 3 tasks, add a brief implementation sketch (approach, key steps, gotchas).\n\n\n\n\n\n- Architecture & design: boundaries, coupling, circular deps, god files, layering violations, scalability bottlenecks\n\n- Code quality: duplication, dead code, complexity hotspots, swallowed exceptions, type-safety holes\n\n- Security: hardcoded secrets, injection risks, missing validation, auth weaknesses, vulnerable dependencies, permissive configs\n\n- Testing: coverage gaps around core logic, tests that assert behavior vs. mere execution, missing test types, untestable code\n\n- Performance: N+1 queries, blocking calls in async paths, missing caching/indexing, unbounded growth\n\n- Dependencies: outdated/unmaintained/heavy packages, license risks, lockfile hygiene\n\n- DevEx & operations: setup friction, CI/CD gaps, lint/format enforcement, logging and observability, deployment story\n\n- Documentation: README accuracy, onboarding path, undocumented critical behavior, docs that contradict code\n\n\n\n\n\nA single document:\n\n1. Executive Summary — ≤10 sentences: overall health grade A–F with justification, top 3 risks, top 3 opportunities.\n\n2. Repo Map — purpose, stack, architecture sketch, key directories with one-line descriptions, anything surprising.\n\n3. Audit Report — findings grouped by dimension, sorted by severity, plus a Strengths section.\n\n4. Improvement Strategy — themes, target states, trade-offs, \"done\" signals.\n\n5. Task Plan — milestones, task table, quick wins, top-3 implementation sketches.\n\n6. Open Questions — anything that needs a human decision (product intent, deprecation candidates, performance targets).\n\n\""},"author":{"name":"Graeme","handle":"gkisokay","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/02a75e36b1e098a2/avatar-33026b4e4f04.jpg","bio":"AI agent enjoyer | @amplifi_now | Helping people scale with AI"},"metadata":{"likes":23,"shares":0,"views":4897,"timestamp":"2026-07-01T20:46:58.000Z"},"media":[],"quotedPost":{"id":"2064685843185647643","platform":"x","url":"https://x.com/gkisokay/status/2064685843185647643","shareMode":"full","content":{"text":"5 Master Prompts to Maximize Fable 5 For Every Build You Have\n\nhttp://x.com/i/article/2053412672176361472"},"author":{"name":"Graeme","handle":"gkisokay","avatar":"https://pbs.twimg.com/profile_images/1988420071824470016/nrXOtFnM_200x200.jpg"},"metadata":{"likes":137,"shares":19,"timestamp":"2026-06-10T12:27:19.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/wxeNp4Txfq/quoted/0.jpg","type":"image","width":1200,"height":480}]},"aiComments":[]},{"id":"2072485418118033701","shareId":"yPSg8vmLyw","platform":"x","url":"https://x.com/Jolyne_AI/status/2072485418118033701","shareMode":"full","createdAt":1783002464337,"postedAt":"2026-07-02T01:00:03.000Z","publishedDate":"2026-07-02T01:00:03.000Z","archivedAt":"2026-07-02T14:27:39.048Z","archivedDate":"2026-07-02T14:27:39.048Z","viewCount":3,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/yPSg8vmLyw","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/vHwYg8jKtz/media/0.png","contentType":"post","content":{"text":"分享一门来自加州大学伯克利分校的进阶课程：Advanced LLM Agents。\n\n这门课聚焦大语言模型 Agent 的最新进展，从推理到规划、从代码到数学证明，系统拆解“能思考、会行动”的代理是怎么做出来的。\n\n课程由 Dawn Song 教授主讲，并邀请 Google、Meta 等公司的研究人员担任客座讲师，内容硬核、案例前沿。\n\n课程地址：https://rdi.berkeley.edu/adv-llm-agents/sp25\n\n你将学到：\n\n- 推理时技术与后训练：把复杂推理能力真正“训出来、用起来”\n\n- 搜索与规划：让代理更会选、更会走、更会做决策\n\n- 代码生成与验证：打造更可靠的编程助手与自动化工具\n\n- 数学定理证明与自动形式化：探索 AI 进入数学的关键路径\n\n- 多模态自主代理：视觉与文本协同，完成更复杂的任务闭环\n\n- AI 安全与安全性：面向部署的可靠性与风险控制\n\n课程提供完整录像、PPT 与配套资料，适合具备机器学习基础的学生和开发者系统进阶。"},"author":{"name":"徐伦","handle":"Jolyne_AI","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/611f857b8affc755/avatar-8a7a176aacfc.jpg"},"metadata":{"likes":271,"shares":49,"views":10156,"timestamp":"2026-07-02T01:00:03.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/vHwYg8jKtz/media/0.png","type":"image","width":417,"height":1247}],"aiComments":[]},{"id":"2072570861761253871","shareId":"923qDLaauo","platform":"x","url":"https://x.com/milesdeutscher/status/2072570861761253871","shareMode":"full","createdAt":1783002453714,"postedAt":"2026-07-02T06:39:34.000Z","publishedDate":"2026-07-02T06:39:34.000Z","archivedAt":"2026-07-02T14:27:29.655Z","archivedDate":"2026-07-02T14:27:29.655Z","viewCount":1,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/923qDLaauo","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072570611021594624/img/J2OXbpp00lwqDAWh.jpg","contentType":"post","content":{"text":"This video will make you a Claude Fable 5 expert in 20 minutes. \n\n0:00 Intro \n\n2:43 wtf is a loop? \n\n5:30 Loop Engineering 101 *demo included* \n\n8:12 Fable 5 x Claude Skills \n\n10:10 3 ways to build effective Fable 5 Skills \n\n12:12 Visual capabilities \n\n14:04 Building a local memory system \n\n18:17 How I'm personally using Fable now"},"author":{"name":"Miles Deutscher","handle":"milesdeutscher","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/b091abdaec61f587/avatar-f027b3042ac8.jpg","bio":"Obsessed with AI. Tweets aren’t financial advice. Building @aiedge_."},"metadata":{"likes":317,"shares":45,"views":47167,"timestamp":"2026-07-02T06:39:34.000Z"},"media":[{"url":"https://video.twimg.com/amplify_video/2072570611021594624/vid/avc1/1280x720/vXgNMV4-8vcqX8Ye.mp4?tag=28","thumbnail":"https://pbs.twimg.com/amplify_video_thumb/2072570611021594624/img/J2OXbpp00lwqDAWh.jpg","type":"video","width":1280,"height":720,"duration":1197}],"aiComments":[]},{"id":"2072253857636180184","shareId":"iMW3dYGoxu","platform":"x","url":"https://x.com/hyeoni0619/status/2072253857636180184","shareMode":"full","createdAt":1782997508184,"postedAt":"2026-07-01T09:39:54.000Z","publishedDate":"2026-07-01T09:39:54.000Z","archivedAt":"2026-07-02T13:05:04.448Z","archivedDate":"2026-07-02T13:05:04.448Z","viewCount":2,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/iMW3dYGoxu","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/mgXF0LFszm/media/0.jpg","contentType":"post","content":{"text":"사람은 '하고 싶은 일'도 \n\n유통기한이 있습니다.\n\n한 미국 여성이 50번째 생일을 맞아\n\n평생 꿈이었던 옐로스톤 국립공원에서 하룻밤을 보내려 했습니다.\n\n젊을 때부터 꼭 해보고 싶었던 일이었지만, 바쁘다는 이유로 계속 미루고 미루다 어느새 50살이 되었죠.\n\n그런데 막상 그 순간이 오자 깨달았다고 합니다.\n\n\"이제는 더 이상 하고 싶지 않다.\"\n\n우리는 하고 싶은 마음이 평생 그대로일 거라고 생각하지만, \n\n사실 꿈에도 '때'가 있는 것 같습니다.\n\n그래서 하고 싶은 일이 생겼다면 '언젠가'보다 '지금'에 조금\n\n 더 가까워지는 게 중요한 것 같아요.\n\n미루다 보면 시간이 없어지는 게 아니라, \n\n내 마음이 먼저 사라질 수도 있으니까요."},"author":{"name":"동글이","handle":"hyeoni0619","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/0042238842315042/avatar-1b858f073bf1.jpg","bio":"사회초년생 성공을 꿈꾸는 20대 주린이 정보글/부업/동기부여/일상🍀"},"metadata":{"likes":1708,"shares":600,"views":180905,"timestamp":"2026-07-01T09:39:54.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/mgXF0LFszm/media/0.jpg","type":"image","width":1079,"height":1339},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/mgXF0LFszm/media/1.jpg","type":"image","width":1078,"height":1370},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/mgXF0LFszm/media/2.jpg","type":"image","width":1071,"height":1360},{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/mgXF0LFszm/media/3.jpg","type":"image","width":1080,"height":1363}],"aiComments":[]},{"id":"DaNy9NfD1U9","shareId":"UU3JT6MZeT","platform":"threads","url":"https://www.threads.com/@samuel369/post/DaNy9NfD1U9","title":"AI Threads","shareMode":"full","createdAt":1782997295866,"postedAt":"2026-06-30T15:46:20.000Z","publishedDate":"2026-06-30T15:46:20.000Z","archivedAt":"2026-07-02T13:01:28.826Z","archivedDate":"2026-07-02T13:01:28.826Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/UU3JT6MZeT","thumbnail":"https://scontent-ssn1-1.cdninstagram.com/v/t51.82787-15/730381890_17933326914321258_7808777332033811146_n.webp?stp=dst-webp_p480x480&_nc_cat=110&ig_cache_key=MzkzMTAyMTIxOTg5Mjk2ODYyMA%3D%3D.3-ccb7-5&ccb=7-5&_nc_sid=58cdad&efg=eyJ2ZW5jb2RlX3RhZyI6IkNBUk9VU0VMX0lURU0ueHBpZHMuMTUzNi5zZHIucmVndWxhcl9waG90by5DMyJ9&_nc_ohc=6wXNBlCsra0Q7kNvwHfw39W&_nc_oc=Adp7e2M94IFWeXl0bECJlgZzxLhgQTkdUTfs06ZAc4DMdFBiqoh1mkPLBMRdLxY7LBM&_nc_ad=z-m&_nc_cid=0&_nc_zt=23&_nc_ht=scontent-ssn1-1.cdninstagram.com&_nc_gid=kWnocvh_9pVlFS2rscqVKA&_nc_ss=7a22e&oh=00_AQB5Dmyv1bX6xPykDJ4yQO7GTIFJ-EOo7Y-cSByEmcfhMg&oe=6A4C1EFB","contentType":"post","content":{"text":"AI Threads\nBeautiful monochrome sexy prompt\nGood night, good rest\nRecreate from\n@mrs.nhette\n번역하기"},"author":{"name":"samuel369","handle":"samuel369","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/91a45cedf5916e53/avatar-f64579148e2f.jpg","bio":"2.7K Followers • 1K Threads • Agar tidak dimanfaatkan org lain, jadilah tidak bermanfaat Prompt enjoyer SPAM LIKES = AUTO BLOCK! See the latest conversations with @samuel369."},"metadata":{"timestamp":"2026-06-30T15:46:20.000Z"},"media":[{"url":"https://scontent-ssn1-1.cdninstagram.com/v/t51.82787-15/730381890_17933326914321258_7808777332033811146_n.webp?stp=dst-webp_p480x480&_nc_cat=110&ig_cache_key=MzkzMTAyMTIxOTg5Mjk2ODYyMA%3D%3D.3-ccb7-5&ccb=7-5&_nc_sid=58cdad&efg=eyJ2ZW5jb2RlX3RhZyI6IkNBUk9VU0VMX0lURU0ueHBpZHMuMTUzNi5zZHIucmVndWxhcl9waG90by5DMyJ9&_nc_ohc=6wXNBlCsra0Q7kNvwHfw39W&_nc_oc=Adp7e2M94IFWeXl0bECJlgZzxLhgQTkdUTfs06ZAc4DMdFBiqoh1mkPLBMRdLxY7LBM&_nc_ad=z-m&_nc_cid=0&_nc_zt=23&_nc_ht=scontent-ssn1-1.cdninstagram.com&_nc_gid=kWnocvh_9pVlFS2rscqVKA&_nc_ss=7a22e&oh=00_AQB5Dmyv1bX6xPykDJ4yQO7GTIFJ-EOo7Y-cSByEmcfhMg&oe=6A4C1EFB","type":"image","width":480,"height":853},{"url":"https://scontent-ssn1-1.cdninstagram.com/v/t51.82787-15/731105034_17933326923321258_5969080958824941154_n.webp?stp=dst-webp_p480x480&_nc_cat=101&ig_cache_key=MzkzMTAyMTIyMDQwNDc1NjkwNw%3D%3D.3-ccb7-5&ccb=7-5&_nc_sid=58cdad&efg=eyJ2ZW5jb2RlX3RhZyI6IkNBUk9VU0VMX0lURU0ueHBpZHMuMTUzNi5zZHIucmVndWxhcl9waG90by5DMyJ9&_nc_ohc=nnSICljRAJgQ7kNvwHYWTvB&_nc_oc=AdpU7CBFpCLaqhGg2QOF8VlZ0wygc4s3FMYh6VmemjkzMRHh893_kC5a3dwuGtaV9iI&_nc_ad=z-m&_nc_cid=0&_nc_zt=23&_nc_ht=scontent-ssn1-1.cdninstagram.com&_nc_gid=kWnocvh_9pVlFS2rscqVKA&_nc_ss=7a22e&oh=00_AQC2xaqh4mZ4_72opX75-EmuzISxUrmi95IaHYiej_jAFg&oe=6A4C47FB","type":"image","width":480,"height":853},{"url":"https://scontent-ssn1-1.cdninstagram.com/v/t51.82787-15/731823749_17933326905321258_4381730025914562892_n.webp?stp=dst-webp_p480x480&_nc_cat=104&ig_cache_key=MzkzMTAyMTIyMTAwODY5NzIyNQ%3D%3D.3-ccb7-5&ccb=7-5&_nc_sid=58cdad&efg=eyJ2ZW5jb2RlX3RhZyI6IkNBUk9VU0VMX0lURU0ueHBpZHMuMTUzNi5zZHIucmVndWxhcl9waG90by5DMyJ9&_nc_ohc=UAf9f6AhSK0Q7kNvwHu23Mh&_nc_oc=Adr-n_dlMhvBgmorJ7oyxq_xIZCJe04zy61eTkKhyOZm9w_dOyCR4ze9YzeOGpPq-XE&_nc_ad=z-m&_nc_cid=0&_nc_zt=23&_nc_ht=scontent-ssn1-1.cdninstagram.com&_nc_gid=kWnocvh_9pVlFS2rscqVKA&_nc_ss=7a22e&oh=00_AQDTlksYROeRzs_RijsAP_HI2iANtnbn7L1WStzZ3WbE2A&oe=6A4C3DF0","type":"image","width":480,"height":853},{"url":"https://scontent-ssn1-1.cdninstagram.com/v/t51.82787-15/731213511_17933326932321258_1629304592020495284_n.webp?stp=dst-webp_p480x480&_nc_cat=104&ig_cache_key=MzkzMTAyMTIyMjcwMzEzNTY3Mw%3D%3D.3-ccb7-5&ccb=7-5&_nc_sid=58cdad&efg=eyJ2ZW5jb2RlX3RhZyI6IkNBUk9VU0VMX0lURU0ueHBpZHMuMTUzNi5zZHIucmVndWxhcl9waG90by5DMyJ9&_nc_ohc=zZTeN9gQsIAQ7kNvwEm1M8P&_nc_oc=Adon3M8dka0BBGW8-r4taDFE4loKHJFE1vY5CNcqPih-WiLLoPpH5S-DyAxtukVRZj4&_nc_ad=z-m&_nc_cid=0&_nc_zt=23&_nc_ht=scontent-ssn1-1.cdninstagram.com&_nc_gid=kWnocvh_9pVlFS2rscqVKA&_nc_ss=7a22e&oh=00_AQCnha29Eq3KCoTNC-QidP9eFVW2ROueRjWepURTY8IqQg&oe=6A4C4387","type":"image","width":480,"height":853}],"comments":[{"id":"DaQYN0GDPGJ","content":"Thx","author":{"name":"hexe_das_boese_ausm_osten","username":"hexe_das_boese_ausm_osten","handle":"hexe_das_boese_ausm_osten","url":"https://www.threads.com/@hexe_das_boese_ausm_osten"}},{"id":"DaOTxPKEwSE","content":"Pretty, thank you\n번역하기","author":{"name":"mrs.nhette","username":"mrs.nhette","handle":"mrs.nhette","url":"https://www.threads.com/@mrs.nhette"}}]},{"id":"2072575193143554498","shareId":"i3nXOe9yK9","platform":"x","url":"https://x.com/sexyguy/status/2072575193143554498","shareMode":"full","createdAt":1782997170739,"postedAt":"2026-07-02T06:56:47.000Z","publishedDate":"2026-07-02T06:56:47.000Z","archivedAt":"2026-07-02T12:59:27.966Z","archivedDate":"2026-07-02T12:59:27.966Z","viewCount":4,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/i3nXOe9yK9","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/d5NdvMaISG/media/0.jpg","contentType":"post","content":{"text":"프롬프트를 3~4번 고쳐쓰는 게 기본값이 됐다면, 이미 돈이랑 시간을 버리고 있는 거다.\n\nGitHub에 별 1만 개 넘게 찍힌 prompt-master라는 Claude 스킬이 있다. 이름만 보면 '또 프롬프트 엔지니어링 강의 같은 거겠지'라고 넘기기 쉬운데, 실제로는 Claude에 직접 설치해서 쓰는 스킬 파일이다. 내가 원하는 결과물을 Claude한테 대충 말하면, Claude가 그걸 해당 AI 도구에 맞는 최적화된 프롬프트로 변환해서 바로 뽑아주는 구조다.\n\n이게 해결하는 문제가 딱 이거다.\n\n막연하게 입력 → 원하는 결과 아님 → 다시 → 또 아님 → 4번째에 겨우 맞음.\n\n나도 매일 이 패턴 반복하는데. API 직접 쓰는 입장에선 그 3번이 다 돈이고, 퇴근 후 남는 시간에 만지는 입장에선 그 시간이 너무 아깝다.\n\n지원 범위가 꽤 넓다. Claude Code, GitHub Copilot, Windsurf, Cursor, Bolt, v0, Lovable 같은 코딩 도구는 물론이고, Midjourney, DALL-E, Stable Diffusion, ComfyUI, Sora, ElevenLabs, Zapier, Make까지 커버한다. 내가 쓰는 AI 도구 거의 다 들어가 있다고 보면 된다.\n\n설치는 간단하다. ZIP 받아서 http://claude.ai → Sidebar → Customize → Skills에 올리면 끝. Claude Code 쓴다면 스킬 디렉토리에 직접 클론해도 된다.\n\n작동 방식이 좀 체계적이다. 내 요청을 task, input, output, constraints, context, audience, memory, success criteria, examples — 9가지 차원으로 분해하고, 빠진 게 있으면 최대 3개만 추가로 물어본다. 그 다음 도구에 맞는 프롬프트 구조로 자동 라우팅하고, 토큰 효율 심사까지 돌린 다음 깔끔한 한 블록으로 결과를 뽑아주는 구조. 3개 이상 안 물어본다는 게 마음에 든다. 어떤 도구들은 맥락 파악한다고 끝없이 물어봐서 오히려 더 피곤하거든.\n\n이 도구의 핵심 문장이 하나 있다: '최고의 프롬프트는 가장 긴 게 아니다. 모든 단어가 결과에 기여하는 게 최고다.'\n\n맞는 말인데... 그럼 나는 지금까지 불필요한 단어로 채운 프롬프트를 몇 백 번은 날린 거라는 얘기잖아. 좀 찔렸다.\n\nClaude Code에서 쓰는 방식이 꽤 명확하다. 복잡한 리팩터링 지시나 아키텍처 리뷰 요청 같은 거 그냥 던지기 전에 prompt-master로 먼저 정제해서 넣으면, 한 번에 원하는 결과 나올 가능성이 확실히 높아진다. Midjourney나 ComfyUI 같은 이미지 도구도 같이 쓰는 사람이면 같은 Claude 환경에서 이미지 프롬프트까지 처리할 수 있으니까 범용성은 있는 편이다.\n\n별 1만이면 그냥 넘기기 좀 애매한 숫자다. 나는 프롬프트 잘 쓴다고 생각했는데, 이걸 보니까 그냥 실패에 익숙해진 거였던 것 같기도 하고.\n\n출처: GitHub — nidhinjs/prompt-master"},"author":{"name":"개아범","handle":"sexyguy","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/860f0b6cb50267ad/avatar-4b2659b58056.jpg"},"metadata":{"likes":68,"shares":15,"views":3224,"timestamp":"2026-07-02T06:56:47.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/d5NdvMaISG/media/0.jpg","type":"image","width":1200,"height":600}],"aiComments":[]},{"id":"2072475339511431345","shareId":"JGryz6dbHm","platform":"x","url":"https://x.com/cevenif/status/2072475339511431345","shareMode":"full","createdAt":1782997153421,"postedAt":"2026-07-02T00:20:00.000Z","publishedDate":"2026-07-02T00:20:00.000Z","archivedAt":"2026-07-02T12:59:09.985Z","archivedDate":"2026-07-02T12:59:09.985Z","viewCount":6,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/JGryz6dbHm","thumbnail":"https://social-archiver-api.social-archive.org/media/archives/yena/QxZi9Fsc0B/media/0.jpg","contentType":"post","content":{"text":"之前 GitHub 热榜简直被 Agent 和 CLI 基础设施承包了，那时的星标增长速度看得人眼花缭乱，不是夸张，是真有这么猛。\n\nGas哥讲几个最炸的项目：\n\n1️⃣ openhuman（tinyhumansai）\n\n私人超级 AI 智能体，Rust 编写，完全私有化部署，日常事务到复杂工作流全包。它不像 ChatGPT 只是个聊天窗口，而是真正跑在你本地的随身助理，连接 Gmail、Notion、GitHub 等 118 个服务后，几分钟就能吃透你的全部上下文。私人 AI 时代不是要来了，是已经来了。\n\n🔗 https://github.com/tinyhumansai/openhuman\n\n2️⃣ CLI-Anything（HKUDS）\n\n口号是\"让所有软件都具备 Agent 原生能力\"，把任意命令行工具变成 AI 能直接调用的技能，让 Agent 真正掌控你的终端和整套工具链。开发者看到这个应该懂含金量。\n\n🔗 https://github.com/HKUDS/CLI-Anything\n\n3️⃣ academic-research-skills（Imbad0202）\n\n专为 Claude Code 打造的学术研究技能包，research→write→review→revise→finalize，整条科研流水线直接自动化跑通。写论文的朋友，这东西都出来了你还在手动搞？\n\n4️⃣ superpowers（obra）\n\n这套不只是工具，而是一套 Agent 技能框架加软件开发方法论，教你怎么系统性地把 Agent 用好。今天的 v6.0 大版本重写了子代理驱动的任务审查流程，在评估中 Claude Code 和 Codex 在相近质量下速度快了约一倍，token 消耗减少近 50%。\n\n🔗 https://github.com/obra/superpowers\n\n5️⃣ agentmemory（rohitg00）\n\nAI 编码 Agent 的持久化记忆方案，基于真实基准测试优化。Agent 最大的硬伤就是没有记忆，长任务跑着跑着就丢了上下文，这个项目专门来补这个坑。编码 Agent 基础设施这一环终于补上了。\n\n🔗 https://github.com/rohitg00/agentmemory\n\n说句实在话，Agent、CLI、Memory、Skills 这四块已经是标配了，不是将来，就是现在。谁先把这套基础设施吃透，下一波 AI 生产力爆发的时候，你就是那个早已准备好的人。"},"author":{"name":"Gas留在链上","handle":"cevenif","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/7c675bae26786dfe/avatar-984c5ae9d0f7.jpg"},"metadata":{"likes":19,"shares":3,"views":1817,"timestamp":"2026-07-02T00:20:00.000Z"},"media":[{"url":"https://social-archiver-api.social-archive.org/media/archives/yena/QxZi9Fsc0B/media/0.jpg","type":"image","width":943,"height":790}],"aiComments":[]},{"id":"2072334422414876957","shareId":"kZqiyqguWY","platform":"x","url":"https://x.com/sdhilip/status/2072334422414876957","shareMode":"full","createdAt":1782997131655,"postedAt":"2026-07-01T15:00:03.000Z","publishedDate":"2026-07-01T15:00:03.000Z","archivedAt":"2026-07-02T12:58:47.970Z","archivedDate":"2026-07-02T12:58:47.970Z","viewCount":5,"ownerUsername":"yena","canonicalUrl":"https://social-archive.org/yena/kZqiyqguWY","contentType":"post","content":{"text":"My Claude Code setup used to be 3 skills.\n\nNow it's 7. Each one earned its spot.\n\ndata engineering (@AnthropicAI) → dbt and Airflow skills that slot right into my data pipeline work.\n\ngstack (@garrytan) → Turns Claude Code into a virtual team that questions what you're building before code starts.\n\ngrill me (@mattpocockuk) → Asks you sharp questions first, so Claude builds what you actually meant.\n\nsuperpowers (obra) → Forces spec, plan, then test-driven development. Solid code, not vibe-coded slop.\n\nimpeccable (@pbakaus) → One command to audit and polish a dashboard or UI until it looks sharp.\n\nlast30days (@mvanhorn) → Finds what's working right now across Reddit, X, YouTube, and HN.\n\nprinting press (@mvanhorn @trevin) → Turns any API or website into a clean, token-light CLI you can actually use.\n\nStart small.\n\nAdd one only when you hit a job your current skills can't do.\n\nNext on my list: compound engineering (@trevin). Still exploring it."},"author":{"name":"Dhilip Subramanian","handle":"sdhilip","avatar":"https://social-archiver-api.social-archive.org/media/authors/yena/1198aac5faf99a38/avatar-2c282636e441.jpg","bio":"Building Data Platforms & AI Solutions | Building https://t.co/wi2QLQBu42 🌍 40/170 countries"},"metadata":{"likes":350,"shares":12,"views":22861,"timestamp":"2026-07-01T15:00:03.000Z"},"media":[],"aiComments":[]}],"total":382,"pagination":{"page":4,"limit":50,"total":382,"hasMore":true}}}

Author posts