Player not loading? Watch on YouTube
Jamie walks through Karpathy's LLM Wiki approach using Obsidian and Claude Code. The idea is to have an AI agent turn source documents into persistent, linked Markdown pages, then consult those pages when answering questions. Jamie contrasts this with a basic RAG workflow that retrieves document chunks for each query; his description should not be read as a universal account of every retrieval system.
The setup starts with an Obsidian vault containing raw and wiki folders. An optional templates folder supports manually created notes. A CLAUDE.md file in the vault root defines the wiki's purpose, ingestion steps, page formatting and citation rules. Obsidian is the viewer; Claude Code reads sources and maintains the pages. The walkthrough uses the Obsidian Web Clipper in Chrome to save articles as Markdown. Jamie also says Claude Code can read PDFs and text files.
Two Japan travel articles provide the demonstration. Claude creates neighborhood pages, updates them with food information, and answers a question about staying near food and temples. A lint request checks the wiki for contradictions, orphan pages and broken links. The report also flags an uningested food source and citation issues, so review remains necessary.
Jamie recommends personal-scale use and cites roughly 100 articles as an example. Larger collections need more infrastructure, and poor sources or AI mistakes can affect the result. The wiki files reside on the user's computer, but the tutorial does not establish offline model execution or private processing.