Claude-Powered Markdown Wikis Beat RAG for Personal Knowledge
Andrej Karpathy's LLM wiki uses Claude to auto-organize raw markdown into linked, indexed notes—setup in 5 minutes, handles 100 docs/500k words, cuts token use 95% vs RAG by reading relationships instead of embeddings.
Setup Claude Wiki in 5 Minutes for Compounding Knowledge
Paste Karpathy's gist prompt into Claude Code (via terminal or VS Code) to initialize a vault: creates /raw for source docs, /wiki for organized output, index.md listing concepts/entities/sources/people/comparisons, log.md for operation history, and claude.md defining project rules. Use free Obsidian as visual frontend for graph view of backlinks/tags. Drop raw content (e.g., PDFs, web clips via Obsidian Web Clipper extension set to /raw) and command Claude: "Ingest file"—it chunks into 5-25 linked MD pages per article, extracts tags/authors/takeaways, builds relationships (e.g., one AI-2027 article yielded 23 pages: 1 source, 6 people, 5 orgs, 1 AI system, multiple concepts). Batch ingest scales: 36 YouTube transcripts in 14 minutes auto-linked tools like Claude Code/WAT framework across videos, revealing patterns without manual work. Customize via claude.md (e.g., flat structure for personal brain vs subfolders like /tools/ /techniques for YouTube wiki). Patiently wait 10-14 minutes per batch as Claude reasons on granularity/focus.
Query and Maintain for Deeper Insights Than Ephemeral Chat
Claude reads full wiki/index/log for queries, following links for context (e.g., click "OpenAI" from source to related model spec/psychology pages). Auto-maintains summaries/index, identifies gaps (e.g., "Fetch articles on compute scaling"), runs "lint" checks for inconsistencies/missing data via web searches/new connections. Add hot.md cache (500 chars recent updates) for agent efficiency. Relationships compound: backlinks connect video techniques to tools, enabling pattern discovery (e.g., MCP servers across 36 videos). Token savings hit 95% on 383 files/100+ transcripts—one user query dropped from massive context to compact wiki reads. Linting ensures scalability; log tracks every update.
Outperforms RAG for Small-Scale: Simpler, Cheaper, Relational
Skip vector DBs/embeddings/chunking—markdown files alone suffice for <500k words/100 docs, as LLM navigates explicit links/index vs similarity search. RAG needs ongoing compute/storage; wiki costs only ingest/query tokens (free infra). Deeper reasoning from relationships ("OpenAI links to governance/geopolitics") beats RAG's shallow chunks. Trade-off: scales poorly beyond small wikis (use RAG for massive corpora). Persists knowledge like "tireless colleague"—integrate via claude.md paths (e.g., executive agent reads /wiki/index/hot cache only if needed, avoiding always-on context bloat). Prompt Claude to build from high-level ideas ("Implement Karpathy's vague gist as my AI research brain"), customizing per use (YouTube vs personal).