A Claude Code plugin (and portable skill) that matches a system you describe against the live Encyclopedia of Agentic Coding Patterns and returns the few patterns, antipatterns, and gaps that actually apply to your system — with the reason each one fires and the trap it warns about.
It prunes ~304 catalog entries down to the 4–8 that matter. The value is the pruning, not the listing: a lazy "here are fifteen loosely-relevant patterns" is a horoscope. This tool states why each pattern applies to what you actually described, and leads with the gaps — the destructive path with no approval policy, the router with no eval gate — because that is where the time gets saved.
It reads the catalog live on every run (the catalog is rewritten continuously), so you always match against the current index rather than stale training data.
- SUGGEST — which patterns fit what you described.
- AUDIT (the headline) — what you're missing or getting wrong, including the sections you didn't think to look at (governance, security, testing).
- On Claude Code, a
SessionStarthook refreshes a local cache of the catalog and prints a one-line banner so the tool is discoverable and works offline.
/plugin marketplace add promptmetrics/agent-patterns
/plugin install agentic-patterns@aipatternbook
Then invoke it explicitly with /agentic-patterns, or just describe an agentic
system and ask "which patterns fit?" / "what am I missing?" — it triggers on its own.
To update later: /plugin marketplace update aipatternbook.
Cowork and claude.ai use plain skills rather than the Claude Code plugin/marketplace system. Install the skill directly:
- Download the folder
plugins/agentic-patterns/skills/agentic-patterns/from this repo and zip it (the zip must containSKILL.mdat its top level). - In claude.ai: Settings → Capabilities → Skills → Upload skill, and select the zip.
- Enable the skill.
The SessionStart banner and offline cache are a Claude Code plugin bonus; on Cowork
the skill simply fetches the catalog live on each run.
Paste a PRD, spec, design doc, or a few sentences describing an agent / LLM / orchestration system, and ask for a pattern check. Example:
"Here's my design: a router agent inspects each request and dispatches to one of 40 sub-agents; each sub-agent can write to our CRM after a preview step. Which patterns apply and what am I missing?"
You'll get a short, ranked list like:
Read: A central router dispatching to 40 write-capable sub-agents, with a preview
step before CRM writes.
## Audit — what's missing or at risk
Approval Policy — gap
- Why it fires here: you preview writes but never state which writes need human
confirmation vs. run unattended — the rule the preview is standing in for.
- Watch for: Approval Fatigue — if every write prompts, reviewers rubber-stamp.
- Read first: https://aipatternbook.com/approval-policy.md
...
| Component | Path | Role |
|---|---|---|
| Marketplace | .claude-plugin/marketplace.json |
Lists the plugin; /plugin marketplace add reads this |
| Plugin manifest | plugins/agentic-patterns/.claude-plugin/plugin.json |
Plugin identity + version |
| Skill | plugins/agentic-patterns/skills/agentic-patterns/SKILL.md |
The matching + pruning logic |
| Hook | plugins/agentic-patterns/hooks/hooks.json |
Runs the refresh script on SessionStart |
| Refresh script | plugins/agentic-patterns/scripts/refresh-catalog.sh |
Caches llms.txt, prints banner, always exits 0 |
The skill prefers a live fetch of https://aipatternbook.com/llms.txt and falls back
to the cache at ~/.cache/aipatternbook/llms.txt when offline. The catalog text is
never vendored into this repo — it is read at runtime and not redistributed.
MIT — see LICENSE.