agent-skills: 24 production-grade workflows that stop your AI from shipping shortcuts

What it is

addyosmani/agent-skills is a collection of 24 Claude Code / multi-agent skills that encode the full software development lifecycle — from spec to observability — in a form any coding agent can follow. By Addy Osmani of the Google Chrome team, sitting at 74k stars with cross-platform support for Cursor, Codex, Copilot, Gemini CLI, and more.

The problem it's solving

Left to their own defaults, coding agents reach for the fastest path: inline implementation, skipped tests, no spec, no review. Fine for a quick prototype; the pattern breaks down in anything that has to survive longer than a demo.

These skills embed the judgment that slows the agent down at the right moments: write the spec first (/spec), break it into tasks (/plan), implement incrementally with test coverage (/build), review for security and performance before shipping (/review), then ship with a proper changelog (/ship). There's also a /webperf command for Lighthouse and Core Web Vitals that's more focused than asking the agent to "optimize performance" cold.

The underlying ideas aren't novel — Hyrum's Law, Chesterton's Fence, trunk-based development — but packaging them as invokable skills means the agent actually uses them, instead of you hoping it will remember.

When it genuinely helps

Drop this in at the start of a greenfield project and use /spec before writing a single line of code. The interview-me skill is legitimately useful: it asks the questions you'd otherwise skip (who's the user, what does done mean, what can break) and surfaces gaps before you've sunk three days into the wrong thing.

/build auto is the most interesting variant — it removes manual confirmation between phases, so the agent flows through implementation → verify → review without waiting for you. Works well on bounded, well-specified tasks; less well when the spec is underspecified (the agent will just proceed and be confidently wrong).

# Universal CLI
npx skills add addyosmani/agent-skills

# Claude Code
/plugin marketplace add addyosmani/agent-skills

The honest caveat

24 skills is a lot. You have to know which slash command to reach for at each moment — and that discipline is still on you. If you run /build without having run /spec and /plan first, you haven't changed much: you're still handing the agent an underspecified prompt and hoping. The skills are guardrails, not autopilot.

The other thing: loading all 24 into context at once eats significant window. Most users end up relying on 3–4 of these day-to-day and ignoring the rest. That's fine — but it means the complete senior-engineer lifecycle pitch is closer to a few really good prompts in practice. Manage your expectations accordingly, and you'll get real value.

Original repo: github.com/addyosmani/agent-skills. Maintained by Addy Osmani.