I spent the past week building a personal knowledge base that the LLM manages for you — and I wanted to share what I learned.

The insight came from Andrej Karpathy's post: most people use LLMs as smart search engines. But you can use them as librarians who own the library.

Here's how it works:
Feed it raw content: URLs, PDFs, images, notes

• /kb-compile turns everything into a structured Obsidian wiki with tagged concept articles and backlinks
• /kb-reflect runs automatically after every compile, reads the index, and writes synthesis articles connecting ideas across sources you never explicitly linked
• /kb-ask navigates the wiki like a human expert (reads index first, picks 3-5 relevant articles, answers with citations) — no RAG, no embeddings infrastructure needed

The part that surprised me most: the synthesis articles. Even with 10-15 sources, the reflection pass surfaced connections I hadn't made explicit and wrote them up as standalone articles. That's qualitatively different from search — search finds what you know is there. This surfaces what you didn't know you knew.

The whole system ships as 9 Claude Code skills (plain markdown files) + a Python search CLI. One setup command. Works with existing Obsidian vaults.

Wrote a full technical walkthrough on my blog: https://lnkd.in/gZ44zecd
GitHub: https://lnkd.in/gfZETjJC