A structured, LLM-maintained knowledge base of AI coding agent plugins, skills, and sub-agents. Built for one purpose: query the wiki to quickly create your own skills or sub-agents.
The wiki catalogs quality-rated plugins from across the ecosystem — Claude Code, Codex, Pi, and platform-agnostic tools. Each plugin page tells you what it is, what it includes, and exactly how to use it. Pattern pages surface the recurring techniques so you can see the shared DNA across plugins. When you ask a question, the wiki synthesizes an answer with citations and files it back — so the knowledge compounds.
The main feature. Ask anything against the wiki:
/query
What you can ask:
- "What are the common elements of skills used for creating a plan?"
- "Which plugins use sub-agent delegation, and how do their approaches differ?"
- "How does compound-engineering-plugin's review loop compare to Anthropic's Code Review?"
- "What's the best way to add spec-driven development to my workflow?"
Answers are synthesized from wiki pages with [[wikilinks]] as citations. At the
end, you're offered the chance to file the synthesis back into the wiki — so your
explorations become permanent reference material.
Plugins are scored on a 5-dimension rubric against observable evidence, not star
counts. The methodology lives in wiki/raw/analysis/evaluation-report.md.
| Dimension | What it scores |
|---|---|
| Engineering Quality (30) | Contributor diversity, issue health, commit cadence |
| Third-Party Validation (25) | Official verification, independent reviews, community depth |
| Usability (20) | Quickstart clarity, documentation, scope discipline |
| Sustainability (15) | Maintenance motivation, organizational backing |
| Distribution (10) | Install counts, ecosystem integration |
Red flags (star inflation, restricted issues, single-maintainer bus factor, failed security scans, scope explosion) are subtracted from the total.
/evaluate <repo-url>
Example: /evaluate https://github.com/some/new-plugin
This fetches the repo, scores it against the rubric, and tells you whether it meets the quality bar. If it scores 65+, you'll be offered the option to ingest it into the wiki.
Add a quality-rated plugin to the wiki:
/ingest <url>
Example: /ingest https://github.com/obra/superpowers
The command fetches the plugin's README, documentation, reviews, and community
discussions, saves everything to wiki/raw/web/, then creates or updates plugin
pages, pattern pages, and cross-references.
/lint
Health-checks the entire wiki: contradictions, stale claims, orphan pages, broken links, pattern deduplication, bidirectional similarity enforcement. Run periodically.
wiki/
llmwiki-schema.md ← canonical conventions (read this first)
index.md ← content catalog
log.md ← chronological record
plugins/ ← one page per plugin (summary, usage, patterns)
patterns/ ← recurring techniques (concept + variations)
comparisons/ ← side-by-side comparisons (query output)
sources/ ← metadata for raw source documents
raw/ ← immutable source documents
references/ ← comprehensive reference files
analysis/ ← evaluation reports
web/ ← fetched from URLs
Built on the LLM Wiki pattern by Andrej Karpathy.
Works with Claude Code (via CLAUDE.md and .claude/commands/) and Pi
(via .pi/skills/). Both layers delegate to the same canonical instructions in
.pi/skills/.