// the find
addyosmani/agent-skills
Production-grade engineering skills for AI coding agents.
A pack of 25 Markdown 'skill' files that encode a software lifecycle (spec, plan, build, test, review, ship) as structured prompts for coding agents, plus adapters that wire them into Claude Code, Cursor, Codex, Gemini CLI, and a dozen other tools. It's for teams who want their AI agent to actually follow a process instead of writing whatever it feels like.
There's a real eval suite (evals/cases + evals/fixtures + scripts/run-evals.js) with CI that checks skills actually fire on the right triggers, not just prompts and a README claiming they work. Each skill has explicit anti-rationalization tables and evidence requirements ('tests passing', not 'seems right'), which is a genuine attempt to close the loophole where an agent claims a step is done without proof. Multi-host support is implemented per-tool rather than faked — separate .claude/commands, .gemini/commands, .codex-plugin directories instead of one generic file reused everywhere and hoping it parses.
This is markdown instructions, not enforcement — nothing stops an agent from ignoring a skill file the same way it can ignore any other system prompt; the only actual code hooks (hooks/*.sh) are a thin layer on top. The maintainers' own issue #361 documents that per-skill `npx skills add --skill X` installs silently drop the shared references/ directory, so the fast-path quick-start produces a degraded skill unless you know to do a whole-repo clone instead. It depends on several independently-versioned third-party CLIs (vercel-labs/skills, Codex plugin CLI, gemini skills installer) it doesn't control, so breakage in any of those tools breaks installation with no visibility from this repo. The eval graders themselves are LLM-judged markdown files (evals/plugin/*/graders/*.md), so 'proof it works' ultimately reduces to another LLM's opinion.