In this video I show you how Andrej Karpathy’s “LLM wiki” idea basically 10x’s what most people are doing with Claude Code—because it turns random files into an actual knowledge system you can query. I’m talking: a simple folder of markdown files that compounds over time instead of disappearing like a normal chat. I demo my own setup where Claude Code grabbed transcripts from 36 recent YouTube videos, organized them into pages, tags, and backlinks, and automatically built relationships between tools, techniques, MCP servers, and frameworks—without me manually wiring anything up. Then I walk you through setting it up in about 5 minutes using Obsidian as the front end (so you can see the graph and navigate nodes). The structure is dead simple: a vault with a /raw folder for source docs and a /wiki folder where Claude Code creates index files, logs, and linked pages. I ingest the AI 2027 article live, show how it can split one source into ~20+ connected pages, and explain how you can plug this wiki into other agents (like my executive assistant) for real context. Finally, I compare this to traditional semantic search RAG. No, it doesn’t “kill” RAG—but for small-to-medium knowledge bases, the markdown + index approach is shockingly effective, cheaper on tokens, and way less infrastructure.