Particle.news

Karpathy’s File-Based LLM Wiki Spreads as Developers Build Obsidian Second Brains

The approach cuts out vector databases by letting the model read a markdown index and linked pages.

Overview

  • Developers are reproducing Andrej Karpathy’s workflow that lets an LLM turn raw documents into an interlinked markdown wiki that Obsidian displays.
  • The setup uses a shared folder with raw and wiki directories, a CLAUDE.md control file, and an index page that guides the model through what to read next.
  • Community builds add practical features such as PDF and YouTube transcript ingestion via converters, duplicate detection and merging, and an auto-updated master index.
  • Early reports cite concrete gains, including 44 wiki pages from about 50 sources, clearer link graphs in Obsidian, and roughly 90% fewer tokens versus prompting on raw files.
  • A veteran user framing the file system as a ready graph database describes strong results at large note counts but flags inbox triage as a hard open problem.