Karpathy's LLM Wiki - Full Beginner Setup Guide¶
Summary¶
A step-by-step video tutorial by Teacher's Tech (Jamie) building an LLM Wiki from scratch using Obsidian and Claude Code, using a "planning a trip to Japan" demo scenario.
Key Takeaways¶
Three-Layer Architecture Explained for Beginners¶
- Raw Sources — PDFs, articles, meeting notes; read-only for the AI; source of truth
- The Wiki — Markdown files the AI creates and maintains: index page, concept pages, entity pages, summary comparisons, all interlinked
- The Schema — Rules document (CLAUDE.md for Claude Code) telling the AI how to structure the wiki, handle new sources, and format everything
Schema File Contents¶
The CLAUDE.md/AGENTS.md schema should specify: - Purpose — what the knowledge base is about (one line to customize) - Folder structure — where raw resources and wiki output live - Ingest workflow — read document → extract key concepts → create/update wiki pages → update index → log changes - Page formatting rules — summary at top, source references, links to related concepts - Question answering behavior — consult wiki first, cite sources, flag uncertainty
Setup Walkthrough¶
- Created
raw/,wiki/,templates/folders in Obsidian vault - Dropped CLAUDE.md schema into project root
- Used Obsidian Web Clipper to save a Tokyo travel blog as markdown into
raw/ - Claude Code ingested the source in ~3 minutes, creating structured wiki pages for neighborhoods, temples, etc.
- Added a second source (Japan food guide); Claude updated existing neighborhood pages rather than just creating new ones — demonstrating the compounding effect
- Graph view showed more nodes and connections after the second ingest
Cross-Source Query Demo¶
Asked: "What neighborhood should I stay at if I want to be close to the best food and still near the major temples?"
Claude pulled from neighborhood pages, food pages, and temple pages — connecting dots spread across completely different sources, citing specific wiki pages. This is fundamentally different from basic RAG.
Linting Demo¶
Prompt: "Please lint the wiki."
Claude returned a report checking: orphan pages, broken links, structural soundness, citation issues. Identified the biggest gap as an uningested food source and offered to fix citation issues.
Use Cases Mentioned¶
- Students/researchers — wiki as you read papers; structured knowledge base instead of highlighted PDFs
- Teachers — feed curriculum documents, PD materials; personal teaching wiki that grows
- Businesses — meeting notes, customer calls, project docs; new team members browse wiki instead of Slack history
- Curious readers — track learning from books, podcasts, articles; personal encyclopedia
Limitations Stated¶
- Best at personal scale (~100 articles); tens of thousands of pages needs more infrastructure
- Garbage in, garbage out — sources must be curated
- Requires a coding agent (Claude Code, Codex, or similar); Obsidian alone doesn't do this
- AI can make mistakes (miscategorization, wrong connections) — linting is essential