Eat the Broccoli 🥦

The quality sweep you know you should run and skip anyway.

A tiered, language-agnostic quality-and-pattern audit for any codebase. It pairs the deterministic tooling you already have (lint, types, tests, dependency CVEs, dead-code, silent-failure scanners) with a learned-pattern hunt — a checklist of the hard failure classes that pass every test and still ship broken: races, stale caches, silent fallbacks, null-masking, deploy-staleness.

It's a Claude Code skill, but the checklist is useful to anyone.

What's in it

• Two dials: depth (quick / standard / deep) × scope (changed / module / repo).
• Deterministic tooling, tool-agnostic — wire in your stack's linter / typechecker / scanner (Python & Rust reference stacks included). The pattern hunt needs no tooling at all.
• The pattern hunt — ~20 structural failure classes across 4 families (state & timing · failure visibility · boundaries & contracts · environment & control flow), each with a "is this by design?" disambiguator so you catch real issues without crying wolf on intentional ones.
• A .broccoli-accept file — record confirmed by-design findings so the next sweep skips them.

Install

Drop SKILL.md into your Claude Code skills directory:

~/.claude/plugins/local/<your-plugin>/skills/eat-the-broccoli/SKILL.md

Then say /eat-the-broccoli (or "eat the broccoli").

Contributing

The pattern catalog is meant to grow across languages and harnesses. Hit a failure class that isn't in the table? Open a PR with a row — Name | ✅ by design if … | ❌ broken if …. Real war stories make the best rows.

Integrations

Works with any tools plus a plain .broccoli-accept file — no dependency. For reference setups that bundle the tooling and make the by-design verdicts compound across a team, see INTEGRATIONS.md.

License

MIT — it's a public good. Eat your vegetables. 🥦