llm-wiki-template

A starter template for a second brain that an LLM writes and maintains for you. You curate sources and ask questions; the LLM compiles and keeps a knowledge base current. Knowledge is compiled once, not re-derived on every query.

It is an Obsidian vault committed to git — there is no app to build. The "code" is a markdown corpus plus a small toolchain for indexing and linting.

Credits

Inspired by and built upon Andrej Karpathy's "LLM wiki" idea. Retrieval is powered by qmd, a local markdown search engine by Tobi ("tobi/qmd").

The idea: two layers

• **raw/** — immutable sources. Notes you import, books, articles, meeting

minutes, talks you attended. Never edited. One folder, raw/authored/, holds your own output — talks you gave, docs and posts you published — so the wiki can synthesize your trajectory as a practitioner, not only your reading.

• **wiki/** — the LLM-maintained knowledge base. Entities, topics, sources,

insights, plans, projects — all written and cross-linked by the LLM. Plus wiki/reflections-log.md, an append-only first-person layer for what actually struck you: the one thing a second brain can never reconstruct from sources.

The LLM ingests from raw/, synthesizes into wiki/, and keeps everything cross-referenced. CLAUDE.md is the schema that tells the LLM exactly how — it is kept under 300 lines on purpose; the longer-form reference lives in docs/CLAUDE-MD-EXTENDED.md.