• •

A Friday in April. A new CMS rule had dropped overnight, 1,590 pages of it, and I was reading the summary on my phone between cases. By lunchtime I had five tabs open: the Federal Register PDF, a KFF explainer, two Health Affairs pieces, and a Substack hot take. I needed to understand the model structure well enough to brief a health system CEO the following week. The tabs would stay open for days. I would read, close, reopen, half remember, go back to page forty seven, lose my place, try again.

This is, more or less, what learning looks like for most of us now. Our fields move faster than our study habits. A practicing clinician in 2026 is asked to stay current on evidence, policy, payment models, AI tooling, operational frameworks, and whatever new thing the health system rolled out last quarter. Academic careers ask for the same on top of original scholarship. Operators ask for it on top of a P&L. The friction is not finding information. The friction is turning dense source material into something you can actually hold in your head.

I want to tell you about a tool I have been using heavily over the last six months, and a small piece of infrastructure I built around it that I think is worth sharing.

NotebookLM Is Better Than You RememberGoogle released NotebookLM quietly in 2023. Most people tried it once, shrugged, and moved on. That was a mistake, and it was partly Google’s fault for underselling it.

If you are new to NotebookLM, a handful of other writers have already produced excellent primers. I especially like Wyndo at AI Maker, Ethan Mollick’s One Useful Thing, and Dan Shipper at Every. Any of them will get you oriented. I want to spend our time on my workflows that have changed how I use NotebookLM, and on the skill I built to exploit it.

NotebookLM is a research workspace that grounds every answer in the sources you upload. You drop in PDFs, text, URLs, video transcripts, or Google Docs, and the model operates only on that corpus. It won’t wander into its training data. It won’t confidently cite a paper that does not exist. When it quotes something, it links back to the exact passage in your uploaded source.

The other pieces are equally useful. Audio Overviews generate two host podcast style conversations about your corpus, tuned by a short prompt you write. Mind Maps surface the concept structure of dense material at a glance. Study Guides produce quiz questions with traceable answers. You can also query your corpus conversationally, which turns a slow skim into a fast interrogation.

A new feature I really like is the Interactive Audio Overviews capability. Midway through the AI hosts’ conversation about your corpus, you can tap a button and literally join the podcast. The two hosts pause, turn to you, and ask what you want to explore. You speak a question out loud. They answer it, grounded in your uploaded sources, then pick up the conversation. It is the single closest thing I have used to having a tutor on demand who has already read your materials. On a long walk through dense policy source material, it is extraordinary.

• •

If you have ever wished you could upload a pile of source material and ask it questions for an hour, that is what NotebookLM is for. It is, in my opinion, the single most underrated learning tool available to a clinician right now.

The Problem With Using It WellThe catch is that NotebookLM only performs as well as the prompts you give it and the resources you populate it with. “Summarize these documents” produces a serviceable but shallow summary. “Generate an audio overview” produces a decent but generic one. The quality of the learning experience is almost entirely a function of how well you direct the model.

And here was my problem. Every time I ingested a new corpus, I found myself re-writing the same prompts from scratch.

“Make me an infographic brief that compares the two main models.”

“ Produce a fifteen minute audio overview for the commute.”

“ Generate a teaching deck outline for grand rounds.”

“ Summarize the contested design choices as a two perspective debate. Write a short explainer video storyboard.”

The prompts were never topic agnostic. They worked best when I named the specific constructs I wanted surfaced from the source material. That meant real writing every time, which meant I ran fewer asset types, which meant I got less out of each corpus.

I started keeping a running doc of my best prompts. Then I started keeping audience specific variants, because I noticed I wanted different framings when I was learning for myself, preparing to teach, briefing a CEO, or explaining a model to my mother, who is a chemical engineer and the sharpest lay reader I know.

At some point the running doc became a spec. And a spec for a reusable workflow is exactly what Claude skills are for.

What the Skill DoesThe skill is called notebooklm-learner-pack. When you invoke it in Claude (on the desktop Claude Co-Work app, in a Claude Chat project, or in Claude Code), you give it three things: a topic, a brand voice, and optional source context. Claude returns a single self contained HTML file that I call a Learner Pack.

Every Learner Pack contains 24 copy paste prompt cards, organized into four asset types:

• Ten infographics, each a different structural pattern (side by side comparison, hub and spoke network, timeline, step by step process, decision tree, anatomy breakdown, tiered hierarchy, metrics dashboard, three panel card set, risk and equity bars)

• Five slide deck outlines (executive summary, teaching deck, conference talk, strategic briefing, operational playbook)

• Five audio overview prompts (rapid brief, standard overview, deep dive expert conversation, debate between two perspectives, narrative case study)

• Four video storyboards (short social explainer, mid length explainer, scenario video, demo walk through)

The Pack renders as a scrollable web page with a sticky filter bar at the top. Two dimensions of filtering. One axis is asset type. The other axis is audience lens: ELI5, Newbie, Clinical, Operator, Finance, Deep Dive. Toggling the audience reframes every prompt in place. The same “compare the two models” card reads differently for a fifth grader, a clinician, a CFO, and a senior peer.

Each card has a copy button. You click, paste into NotebookLM, and the Pack does the prompt writing for you.

The prompts themselves are grounded. The skill is instructed to name the specific constructs, actors, and numbers from your source material rather than using generic scaffolds. A CJR-X Learner Pack references MS-DRGs 469 and 470, the 21 HCC flags, the 36 benchmark cells, the three waivers. A quantum mechanics Learner Pack would reference Hilbert spaces, commutators, the specific interpretations covered in your textbook. Generic Packs are failed Packs.

Screenshots from a generated Learner Pack below:

• •

The HTML is fully interactive, you can click your audience type and asset type. The prompts to put into NotebookLM are ready and a copy button is present making a split tab workflow very easy.

• •

Looking for more AI workflows like this? Check out the Techy Flows page. Here is another article walking through how I make grand rounds presentations with a flow from open evidence to claude to gemini.

How I Built ItI built the skill inside a normal Claude session using the skill-creator skill, which is Anthropic’s meta skill for creating other skills. The process was less rigorous than the formal evaluation loop that skill-creator supports, and more like pair programming with a patient collaborator.

I iterated on a single worked example, a CJR-X Learner Pack rendered in my own Techy Surgeon voice, over the course of an afternoon. I would ask Claude to generate a version, open the HTML in my browser, click around, tell Claude what was missing, and iterate again. Each round the pattern got sharper. I added the audience toggle. I added the asset type filter. I added the ELI5 lens after my wife asked what CJR-X was and I realized every explanation I had written was too dense for a smart generalist.

Once the pattern felt right, the abstraction was easy. The HTML is data driven: 24 prompt scaffolds, a table of audience modifiers, a table of brand voices. Claude composes the full prompt text on the fly when you click a toggle. Which means the skill can produce Learner Packs in any brand voice (Techy Surgeon, Duke Health, RevelAi Modern Enterprise, or a custom brand you specify) without re-writing 144 prompt variants by hand.

I think skill building itself is an underrated meta capability for clinicians. If you find yourself re-writing the same workflow a third time, you have found something worth codifying. The codification is not just efficiency. It is a forcing function to make your own tacit expertise explicit, which is a learning exercise in itself.

Subscribe now

• •

Three Workflows I Use That You Can TooHealth PolicyA new CMS model is released. I pull the Federal Register rule, one or two KFF explainers, a handful of Health Affairs pieces, and any legislative history that matters. I drop all of that into a fresh NotebookLM notebook. Then I come to Claude and say: “Make me a Techy Surgeon branded Learner Pack on [model name], here are the documents I uploaded.” I get the Pack in a few minutes.

On the first pass I toggle to Deep Dive and run the Debate audio prompt. That gives me a twenty minute two perspective overview I listen to on my drive home. On day two I toggle to Operator and run three infographic prompts to generate briefs I can feed into Napkin. By day three I am teaching a Grand Rounds on the model from the Teaching Deck outline.

Clinical Topic MasteryI want to come up to speed on a surgical technique or pharmacology topic. OpenEvidence pulls the curated synthesis for me. I add two or three recent journal articles, a review article if one exists, and any IFU documents for a specific device. Fifteen documents total, usually.

I generate a Clinical lens Learner Pack. The Anatomy Breakdown infographic gives me a one page structural overview. The Standard Overview audio is a fifteen minute lecture on the topic I can listen to at 1.5x. The Case Study audio walks me through one concrete scenario, which is how my brain consolidates procedural knowledge. The Conference Talk deck scaffold is ready if someone asks me to lecture.

Deep Topic Mastery Outside MedicineI am working through quantum mechanics this year (just a passing interest :-) ). I downloaded a textbook PDF, added a couple of Scott Aaronson blog posts, and a handful of review articles on interpretations. The corpus is dense and I am mostly a tourist.

I generate an ELI5 and Newbie lens Learner Pack. The ELI5 Audio Overview gave me a twelve minute primer I replayed three times over a week until the vocabulary stuck. The Decision Tree infographic helped me map the interpretations against the experimental phenomena they try to explain. I am working my way up to the Deep Dive lens. The Pack is the spine of a multi week study plan, and I do not have to write a single prompt to move through it.

Conference Abstract SynthesisThis one is optional, and honestly the best surprise. Before AAOS or AOA I now compile every abstract from the sessions I care about into a single notebook. The Standard Overview audio surfaces themes I would have missed. The Three Panel Card infographic lets me spot which three studies cluster around a single question. I walk into the meeting with a map.

How to Install and Use the SkillOne. Download the .skill file from the link at the bottom of this article.

Two. Drop it into your Claude skills directory. On Claude Desktop and Claude Code, the path is typically .claude/skills/. On Cowork, install via the plugin interface.

Three. In a new Claude session, ask for a Learner Pack the way you would ask for anything else: “Make me a Duke branded Learner Pack on hip fracture policy. Here is the Federal Register rule, the ASHA response, and the Hospital for Special Surgery commentary.”

Four. Open the HTML Claude generates, click through the toggles, and start running prompts against your NotebookLM notebook through simple copy paste.

The skill will ask one or two clarifying questions if your inputs are ambiguous. Give it real source material if you have it, or just a topic if you do not. Grounded Packs are better, but topic only Packs still beat generic prompts.

** In this flow and all other flows for NotebookLM you still must curate the sources. There are great ways to find sources embedded right within NotebookLM in their web search, but you need to bring the. material you are interested in learning into NotebookLM, and in this flow you have to copy paste and generate them from the artifact into NotebookLM itself.

• •

The Bigger IdeaEvery clinician I respect is a lifelong learner in some serious and self directed way. What has changed recently is that the tools for learning have become good enough to compound the effort of a motivated person dramatically. NotebookLM is one of those tools. A well designed prompt library on top of it is another. A skill that generates that library on demand for any topic you encounter is a third.

This Learner Pack skill is a small thing. It is also, in a sense, the kind of small thing I want more of: workflows codified, tacit expertise made explicit, the slow habits of a studious clinician turned into something reusable. I would love to see colleagues build their own versions, fork this one, disagree with the taxonomy, add a sixth asset type I did not think of.

If you try it, tell me what breaks. Tell me which workflow you added. Tell me which audience lens you found you actually use. I read every reply.

Subscribe now

Download the skill: Here (Github Repo)

Or clone from URL https://github.com/DrMultivac/notebooklm-learner-pack.git

The skill works on Claude Desktop, Claude Code, and Cowork. Source documents do not leave your NotebookLM notebook. The skill itself only generates prompts; the learning still happens in your browser and in your head, which is where it belongs.

One more thing, for the Claude Code users.There is a NotebookLM MCP server you can connect to Claude Code (link here) that eliminates the copy-paste step entirely. Type “Create a Mind Map for me on CJR-X, clinical lens” in your Claude Code session. Claude calls the NotebookLM MCP, and the Mind Map appears inside your NotebookLM notebook. You never leave the terminal. Same for the Audio Overview. Same for the Flashcards. One session, zero copy-paste. You can have Claude load the Learner Pack skill, composes ALL the prompts, and give it the green light to go generate all the assets you have in mind. Mind blown. One thing I would say is that the notebook learner pack has overall too many resource options and creating all of the resources at once can make it overwhelming— for this auto creation you may want to pare down, or prompt your claude instance to have branding applied.

Video coming soon!

Don’t want to miss the next tutorial? Consider signing up for Techy Surgeon a free or paid subscriber.

• •

A Friday in April. A new CMS rule had dropped overnight, 1,590 pages of it, and I was reading the summary on my phone between cases. By lunchtime I had five tabs open: the Federal Register PDF, a KFF explainer, two Health Affairs pieces, and a Substack hot take. I needed to understand the model structure well enough to brief a health system CEO the following week. The tabs would stay open for days. I would read, close, reopen, half remember, go back to page forty seven, lose my place, try again.

This is, more or less, what learning looks like for most of us now. Our fields move faster than our study habits. A practicing clinician in 2026 is asked to stay current on evidence, policy, payment models, AI tooling, operational frameworks, and whatever new thing the health system rolled out last quarter. Academic careers ask for the same on top of original scholarship. Operators ask for it on top of a P&L. The friction is not finding information. The friction is turning dense source material into something you can actually hold in your head.

I want to tell you about a tool I have been using heavily over the last six months, and a small piece of infrastructure I built around it that I think is worth sharing.

NotebookLM Is Better Than You Remember

Google released NotebookLM quietly in 2023. Most people tried it once, shrugged, and moved on. That was a mistake, and it was partly Google’s fault for underselling it.

If you are new to NotebookLM, a handful of other writers have already produced excellent primers. I especially like Wyndo at AI Maker, Ethan Mollick’s One Useful Thing, and Dan Shipper at Every. Any of them will get you oriented. I want to spend our time on my workflows that have changed how I use NotebookLM, and on the skill I built to exploit it.

NotebookLM is a research workspace that grounds every answer in the sources you upload. You drop in PDFs, text, URLs, video transcripts, or Google Docs, and the model operates only on that corpus. It won’t wander into its training data. It won’t confidently cite a paper that does not exist. When it quotes something, it links back to the exact passage in your uploaded source.

The other pieces are equally useful. Audio Overviews generate two host podcast style conversations about your corpus, tuned by a short prompt you write. Mind Maps surface the concept structure of dense material at a glance. Study Guides produce quiz questions with traceable answers. You can also query your corpus conversationally, which turns a slow skim into a fast interrogation.

A new feature I really like is the Interactive Audio Overviews capability. Midway through the AI hosts’ conversation about your corpus, you can tap a button and literally join the podcast. The two hosts pause, turn to you, and ask what you want to explore. You speak a question out loud. They answer it, grounded in your uploaded sources, then pick up the conversation. It is the single closest thing I have used to having a tutor on demand who has already read your materials. On a long walk through dense policy source material, it is extraordinary.

If you have ever wished you could upload a pile of source material and ask it questions for an hour, that is what NotebookLM is for. It is, in my opinion, the single most underrated learning tool available to a clinician right now.

The Problem With Using It Well

The catch is that NotebookLM only performs as well as the prompts you give it and the resources you populate it with. “Summarize these documents” produces a serviceable but shallow summary. “Generate an audio overview” produces a decent but generic one. The quality of the learning experience is almost entirely a function of how well you direct the model.

And here was my problem. Every time I ingested a new corpus, I found myself re-writing the same prompts from scratch.

“Make me an infographic brief that compares the two main models.”

“ Produce a fifteen minute audio overview for the commute.”

“ Generate a teaching deck outline for grand rounds.”

“ Summarize the contested design choices as a two perspective debate. Write a short explainer video storyboard.”

The prompts were never topic agnostic. They worked best when I named the specific constructs I wanted surfaced from the source material. That meant real writing every time, which meant I ran fewer asset types, which meant I got less out of each corpus.

I started keeping a running doc of my best prompts. Then I started keeping audience specific variants, because I noticed I wanted different framings when I was learning for myself, preparing to teach, briefing a CEO, or explaining a model to my mother, who is a chemical engineer and the sharpest lay reader I know.

At some point the running doc became a spec. And a spec for a reusable workflow is exactly what Claude skills are for.

What the Skill Does

The skill is called notebooklm-learner-pack. When you invoke it in Claude (on the desktop Claude Co-Work app, in a Claude Chat project, or in Claude Code), you give it three things: a topic, a brand voice, and optional source context. Claude returns a single self contained HTML file that I call a Learner Pack.

Every Learner Pack contains 24 copy paste prompt cards, organized into four asset types:

• Ten infographics, each a different structural pattern (side by side comparison, hub and spoke network, timeline, step by step process, decision tree, anatomy breakdown, tiered hierarchy, metrics dashboard, three panel card set, risk and equity bars)

• Five slide deck outlines (executive summary, teaching deck, conference talk, strategic briefing, operational playbook)

• Five audio overview prompts (rapid brief, standard overview, deep dive expert conversation, debate between two perspectives, narrative case study)

• Four video storyboards (short social explainer, mid length explainer, scenario video, demo walk through)

The Pack renders as a scrollable web page with a sticky filter bar at the top. Two dimensions of filtering. One axis is asset type. The other axis is audience lens: ELI5, Newbie, Clinical, Operator, Finance, Deep Dive. Toggling the audience reframes every prompt in place. The same “compare the two models” card reads differently for a fifth grader, a clinician, a CFO, and a senior peer.

Each card has a copy button. You click, paste into NotebookLM, and the Pack does the prompt writing for you.

The prompts themselves are grounded. The skill is instructed to name the specific constructs, actors, and numbers from your source material rather than using generic scaffolds. A CJR-X Learner Pack references MS-DRGs 469 and 470, the 21 HCC flags, the 36 benchmark cells, the three waivers. A quantum mechanics Learner Pack would reference Hilbert spaces, commutators, the specific interpretations covered in your textbook. Generic Packs are failed Packs.

Screenshots from a generated Learner Pack below:

The HTML is fully interactive, you can click your audience type and asset type. The prompts to put into NotebookLM are ready and a copy button is present making a split tab workflow very easy.

Looking for more AI workflows like this? Check out the Techy Flows page. Here is another article walking through how I make grand rounds presentations with a flow from open evidence to claude to gemini.

How I Built It

I built the skill inside a normal Claude session using the skill-creator skill, which is Anthropic’s meta skill for creating other skills. The process was less rigorous than the formal evaluation loop that skill-creator supports, and more like pair programming with a patient collaborator.

I iterated on a single worked example, a CJR-X Learner Pack rendered in my own Techy Surgeon voice, over the course of an afternoon. I would ask Claude to generate a version, open the HTML in my browser, click around, tell Claude what was missing, and iterate again. Each round the pattern got sharper. I added the audience toggle. I added the asset type filter. I added the ELI5 lens after my wife asked what CJR-X was and I realized every explanation I had written was too dense for a smart generalist.

Once the pattern felt right, the abstraction was easy. The HTML is data driven: 24 prompt scaffolds, a table of audience modifiers, a table of brand voices. Claude composes the full prompt text on the fly when you click a toggle. Which means the skill can produce Learner Packs in any brand voice (Techy Surgeon, Duke Health, RevelAi Modern Enterprise, or a custom brand you specify) without re-writing 144 prompt variants by hand.

I think skill building itself is an underrated meta capability for clinicians. If you find yourself re-writing the same workflow a third time, you have found something worth codifying. The codification is not just efficiency. It is a forcing function to make your own tacit expertise explicit, which is a learning exercise in itself.

Subscribe now

Three Workflows I Use That You Can Too

Health Policy

A new CMS model is released. I pull the Federal Register rule, one or two KFF explainers, a handful of Health Affairs pieces, and any legislative history that matters. I drop all of that into a fresh NotebookLM notebook. Then I come to Claude and say: “Make me a Techy Surgeon branded Learner Pack on [model name], here are the documents I uploaded.” I get the Pack in a few minutes.

On the first pass I toggle to Deep Dive and run the Debate audio prompt. That gives me a twenty minute two perspective overview I listen to on my drive home. On day two I toggle to Operator and run three infographic prompts to generate briefs I can feed into Napkin. By day three I am teaching a Grand Rounds on the model from the Teaching Deck outline.

Clinical Topic Mastery

I want to come up to speed on a surgical technique or pharmacology topic. OpenEvidence pulls the curated synthesis for me. I add two or three recent journal articles, a review article if one exists, and any IFU documents for a specific device. Fifteen documents total, usually.

I generate a Clinical lens Learner Pack. The Anatomy Breakdown infographic gives me a one page structural overview. The Standard Overview audio is a fifteen minute lecture on the topic I can listen to at 1.5x. The Case Study audio walks me through one concrete scenario, which is how my brain consolidates procedural knowledge. The Conference Talk deck scaffold is ready if someone asks me to lecture.

Deep Topic Mastery Outside Medicine

I am working through quantum mechanics this year (just a passing interest :-) ). I downloaded a textbook PDF, added a couple of Scott Aaronson blog posts, and a handful of review articles on interpretations. The corpus is dense and I am mostly a tourist.

I generate an ELI5 and Newbie lens Learner Pack. The ELI5 Audio Overview gave me a twelve minute primer I replayed three times over a week until the vocabulary stuck. The Decision Tree infographic helped me map the interpretations against the experimental phenomena they try to explain. I am working my way up to the Deep Dive lens. The Pack is the spine of a multi week study plan, and I do not have to write a single prompt to move through it.

Conference Abstract Synthesis

This one is optional, and honestly the best surprise. Before AAOS or AOA I now compile every abstract from the sessions I care about into a single notebook. The Standard Overview audio surfaces themes I would have missed. The Three Panel Card infographic lets me spot which three studies cluster around a single question. I walk into the meeting with a map.

How to Install and Use the Skill

One. Download the .skill file from the link at the bottom of this article.

Two. Drop it into your Claude skills directory. On Claude Desktop and Claude Code, the path is typically .claude/skills/. On Cowork, install via the plugin interface.

Three. In a new Claude session, ask for a Learner Pack the way you would ask for anything else: “Make me a Duke branded Learner Pack on hip fracture policy. Here is the Federal Register rule, the ASHA response, and the Hospital for Special Surgery commentary.”

Four. Open the HTML Claude generates, click through the toggles, and start running prompts against your NotebookLM notebook through simple copy paste.

The skill will ask one or two clarifying questions if your inputs are ambiguous. Give it real source material if you have it, or just a topic if you do not. Grounded Packs are better, but topic only Packs still beat generic prompts.

** In this flow and all other flows for NotebookLM you still must curate the sources. There are great ways to find sources embedded right within NotebookLM in their web search, but you need to bring the. material you are interested in learning into NotebookLM, and in this flow you have to copy paste and generate them from the artifact into NotebookLM itself.

The Bigger Idea

Every clinician I respect is a lifelong learner in some serious and self directed way. What has changed recently is that the tools for learning have become good enough to compound the effort of a motivated person dramatically. NotebookLM is one of those tools. A well designed prompt library on top of it is another. A skill that generates that library on demand for any topic you encounter is a third.

This Learner Pack skill is a small thing. It is also, in a sense, the kind of small thing I want more of: workflows codified, tacit expertise made explicit, the slow habits of a studious clinician turned into something reusable. I would love to see colleagues build their own versions, fork this one, disagree with the taxonomy, add a sixth asset type I did not think of.

If you try it, tell me what breaks. Tell me which workflow you added. Tell me which audience lens you found you actually use. I read every reply.

Subscribe now

Download the skill: Here (Github Repo)

Or clone from URL https://github.com/DrMultivac/notebooklm-learner-pack.git

The skill works on Claude Desktop, Claude Code, and Cowork. Source documents do not leave your NotebookLM notebook. The skill itself only generates prompts; the learning still happens in your browser and in your head, which is where it belongs.

One more thing, for the Claude Code users.

There is a NotebookLM MCP server you can connect to Claude Code (link here) that eliminates the copy-paste step entirely. Type “Create a Mind Map for me on CJR-X, clinical lens” in your Claude Code session. Claude calls the NotebookLM MCP, and the Mind Map appears inside your NotebookLM notebook. You never leave the terminal. Same for the Audio Overview. Same for the Flashcards. One session, zero copy-paste. You can have Claude load the Learner Pack skill, composes ALL the prompts, and give it the green light to go generate all the assets you have in mind. Mind blown. One thing I would say is that the notebook learner pack has overall too many resource options and creating all of the resources at once can make it overwhelming— for this auto creation you may want to pare down, or prompt your claude instance to have branding applied.

Video coming soon!