In this video, I tested Andrej Karpathy’s new “LLM Wiki” idea — basically a self-evolving knowledge system where the model does the annoying part for you: it reads raw data, writes summaries, builds structure, maintains consistency, and keeps improving over time. The big shift here is Karpathy didn’t ship a polished app — he shipped an “idea file” blueprint you can hand to an agent (like Claude Code) and have it implement the whole system for your exact workflow. And when you hook that into a coding agent, it unlocks a totally different level of code generation because your agent can navigate a structured wiki instead of relying on fragile chat memory. I walk you through setting it up fast using Obsidian as the front-end: create a vault, open it in Claude Code (I use it inside VS Code), paste Karpathy’s gist, and give the agent a better prompt so it generates the folder structure, index, and scripts to ingest new sources. The architecture is simple: Raw Sources (your untouched truth), Wiki (LLM-generated markdown with links), and Schema Rules (how the model maintains everything). Then you clip docs/pages into RAW (Obsidian Web Clipper is clutch), tell Claude to compile them into the wiki, and periodically run a “lint/health check” prompt to fix contradictions, stale info, missing links, and new connections. The result is a self-healing memory system that makes Claude Code way more accurate, less hallucination-prone, and way more useful for real workflows like UI/landing page generation.