llm-wiki-skill

Why we included this project

This is a working implementation of Andrej Karpathy's llm-wiki idea. Instead of re-retrieving raw documents on every query, the agent compiles sources into a persistent, cross-linked Markdown wiki, and each ingest, query, lint, and audit pass makes that wiki richer. The workflow is simple: drop markdown sources into a folder, tell the agent to ingest them, then ask questions against the compiled wiki. That suits a research deep-dive, a personal encyclopedia, or a team knowledge base built from meeting notes and Slack threads. The repo also includes real tooling, not just a prompt: a local Node.js preview server that renders mermaid, KaTeX, and wikilinks, plus an Obsidian plugin for leaving severity-tagged comments on any page. Both write byte-identical audit files through a shared TypeScript library, so feedback filed in the browser or in Obsidian lands in the same review loop. The project is explicitly experimental and small, but the division of labor is clearly documented: you source material and steer, the LLM handles writing and cross-referencing. That makes it easy to evaluate on your own notes.