2,000 Notes Nobody Reads—Karpathy's Rule Already Called It

2,000 notes are sitting in Obsidian, folders inside folders, 40 tabs still open, 6 PDFs half-read. When writing them, you thought you'd open them again; you didn't. On X, mojesko calls this kind of library a graveyard and traces the cause to a rule from Andrej Karpathy: if you can't build it from scratch, you don't really understand it. The original post puts it bluntly: tutorials are fake learning; your brain deletes them after three days. Karpathy is a founding member of OpenAI, former director of AI at Tesla, and has long taught building from scratch. The rule is unpleasant to hear, but it explains why saving something isn't the same as understanding it. Obsidian is a local Markdown notes app; a vault is that local folder. Claude Code is Anthropic's terminal coding agent, able to read local files and execute commands. mojesko's approach is to point Claude Code at the vault folder. 5,000 notes become a queryable body of material. Ask a question, and the answer stays as close as possible to your own notes' wording and evidence, with less model guesswork. This claim has a precondition. If the notes are low quality and the links are messy, retrieval just returns a pile of old noise. The loop closes in another example. If you want to learn neural networks, don't start by opening a 3-hour video. Have Claude Code write a small 200-line network from scratch in the vault, watch it train, deliberately break a layer, watch it fail, then fix it and understand why it failed. The original post says you can break through in 20 minutes, replacing 3 weeks. Notes come from feedback. One idea per file, link it to 10 notes, and write it down while the memory is still fresh. That's how compounding happens: 60 notes in month 1, 900 by month 6. New notes pull in old notes. Next time you ask a question, more of the answer comes from your own material and less from Reddit and other people's threads. Setup is described as taking 4 minutes: plain text, no lock-in, no paid plugins. The original post includes a screen recording of the interface, where you can see the network graph, chat window, file inspection panel, and GRAPH INDEX.

Some commenters added the other side in replies. Alan K says his notes are re-read by the agent at the start of every session. The files only grow, so every load gets more expensive. His logs and decision records roll by date; on Sundays he archives the previous week, and the cloud agent reads at most about 1 MB at a time. The real cost is loading; setup only takes four minutes. karter's reply is more direct: storage does not equal knowledge. If you can't rebuild it, the library is just a prettier graveyard. The real value of this approach is turning learning from collecting into building. The idea that it makes your notes automatically smarter is a misunderstanding. Claude Code can't understand for you, but it can force you to expose what you don't understand. Anyone who wants to try it can do so over the weekend. Don't start by adding 2,000 notes; first have Claude Code answer one question from your old notes, then have it build a small thing from scratch. If you can't build it, your notes are still just a bookmark folder. Original post: https://x.com/mojeskoqq/status/2116487711553102046