llm-wiki

A Claude Code plugin that builds and maintains an Open Knowledge Format (OKF) knowledge base under a .llm-wiki/ directory: an AI-native (index-first, no vector DB), human-readable (plain markdown), git-shareable, and recursively composable wiki.

• OKF (spec) — markdown + YAML frontmatter, organized for progressive disclosure.
• Karpathy's "LLM Wiki" (source) — the LLM incrementally builds/maintains a

persistent, interlinked wiki that sits between you and raw sources.

Architecture: thin skills + one deterministic core

The plugin is seven thin skills over one tested deterministic core (llm-wiki-core, Python). The model decides meaning (relevance, ontology, synthesis); the core decides mechanics (identity, hashing, discovery, serialization, mount/dedupe, link integrity, atomic writes, secret scanning). See references/DESIGN-PLAN.md for the full design (Rev 6) and references/research/evidence-matrix.md for the verified research basis.

Install (local directory marketplace)

This repo is a self-hosting single-plugin marketplace. To use the /llm-wiki:* commands, register it once:

Then restart Claude Code so the skills load. (Registration mutates your Claude config; do it yourself when ready.)

Requirements

• Python 3 (stdlib + pyyaml) for llm-wiki-core.
• Optional, for secret scanning: detect-secrets and/or gitleaks (pinned). If absent, scans warn and block promotion

rather than passing silently.