A public LLM Wiki, an AI-maintained, interlinked markdown knowledge base built with the LLM Wiki pattern by Andrej Karpathy. It serves two purposes at once:
- The public knowledge base for sloop: the local-first, cross-repo agent-fleet tool. Domain, architecture, and design concepts, summarized and cross-linked.
- A worked example you can fork. The whole thing is the demo: a real vault, built by
prompting an AI to read sources and file them into a compounding wiki. See
PLAYBOOK.mdfor the exact prompt-by-prompt walkthrough.
Classic RAG re-reads raw documents on every question. An LLM Wiki instead has the AI
incrementally build and maintain a persistent wiki: source summaries, entity pages,
concept pages, overviews, all interlinked with [[wikilinks]]. Knowledge compounds; the
tedious bookkeeping (summaries, cross-references, index, changelog) is exactly what the AI is
there to do, so the wiki actually stays maintained.
| Path | What |
|---|---|
AGENTS.md |
Canonical rules & schema, read this first (tool-agnostic) |
PLAYBOOK.md |
How this vault was built, prompt by prompt |
index.md |
The map, catalog of every page by category |
log.md |
Append-only changelog of ingests / queries / lints |
raw/ |
Immutable source materials (the AI reads, never edits) |
wiki/ |
AI-maintained pages: sources/, entities/, concepts/, overviews/, projects/ |
_templates/ |
Page templates |
CLAUDE.md, GEMINI.md |
Thin pointers so those CLIs load AGENTS.md |
- Browse in Obsidian (open this folder as a vault) for graph view + backlinks, or just read the markdown on GitHub.
- Extend it with any AI CLI that reads
AGENTS.md(Claude Code, Codex, Cursor, Gemini, Copilot, Antigravity). Point it at a source and say "ingest this." - Publish it as a website: Quartz / Obsidian Publish / MkDocs (publisher TBD).
Content is intended to be freely readable and forkable. sloop itself is MIT.