@NickSpisak_ - Karpathy's Second Brain Setup

Open original

@NickSpisak_ on X This is a how-to thread built on Andrej Karpathy's idea of keeping a personal knowledge base as plain markdown files that an AI compiles into a wiki. It walks through a three-folder structure, a CLAUDE.md schema file, and Vercel's agent-browser tool for automatically pulling web pages into the raw sources folder.

  • The setup needs only three folders (raw, wiki, outputs) and one schema file; no database, plugins, or paid apps are required.
  • Vercel Labs' agent-browser, a free CLI tool, lets an AI agent scrape JavaScript-heavy or login-gated pages straight into the raw folder, and the author claims it uses 82% fewer tokens than Playwright MCP.
  • The author's main warning is that errors compound: AI-generated answers filed back into the wiki can propagate mistakes, so periodic health checks are needed.

Karpathy's idea, as relayed here, is to stop scattering notes across apps and instead dump everything into one plain folder of text and markdown files. From there, an AI does the organizing: it reads the raw material, builds a wiki of summaries and cross-links, and maintains an index. The setup described is three folders (raw/, wiki/, outputs/), with raw/ as an unsorted junk drawer you never tidy by hand. The author adds an optional step using Vercel Labs' agent-browser, a free CLI that lets an AI agent control a real Chrome instance to scrape pages that copy-paste can't handle, like JavaScript-heavy sites or content behind logins. He claims it uses 82% fewer tokens than Playwright MCP, which is the sort of figure worth checking yourself before relying on it. The part that makes the system work, according to the author, is a single schema file, usually named CLAUDE.md or AGENTS.md, that tells the AI what the knowledge base is for and how to structure it. Karpathy reportedly keeps his "super simple and flat," and the author's own version is just a text file of rules. Once the wiki exists, the workflow is to ask questions against it, such as identifying gaps, comparing what two sources say, or drafting a briefing from only the collected material, and then save good answers back into the knowledge base so later answers build on them. The one real risk he flags, raised by a commenter, is that errors compound when outputs get filed back in. His fix is a periodic health check where the AI looks for contradictions, unexplained topics, and claims with no source in raw/. The article also pushes back on the Obsidian-plugin crowd: a folder of markdown files and a good schema, he argues, beats a heavily configured tool stack most of the time. Worth noting that the piece is also a funnel for his Twitter account and a free "LLM Knowledge Base" skill, so the setup advice is practical but the framing is promotional. The core method, though, doesn't depend on buying anything and is easy to try in an afternoon.

Ask about this article