agent-skills is Addy Osmani’s collection of production-grade engineering skills for AI coding agents. The framing is specific: not prompts that make an agent “better,” but encoded workflows and quality gates that senior engineers use, organized across the software lifecycle from defining work to shipping it. The author matters here. Osmani is a Google engineering director and a long-time web-performance authority, and the skills bake in concepts from Google’s own engineering culture. That pedigree is both the reason it drew tens of thousands of stars quickly and the lens for judging it.
What is in it
There are 23 skills (22 lifecycle skills plus one meta-skill that handles routing), grouped by phase:
- Define: interview-me, idea-refine, spec-driven-development.
- Plan: planning-and-task-breakdown.
- Build: incremental-implementation, test-driven-development, context-engineering, source-driven-development, frontend-ui-engineering, api-and-interface-design, and more.
- Verify: browser-testing-with-devtools, debugging-and-error-recovery.
- Review: code-review-and-quality, code-simplification, security-and-hardening, performance-optimization.
- Ship: git-workflow-and-versioning, ci-cd-and-automation, documentation-and-adrs, shipping-and-launch.
Each skill is a structured SKILL.md with a process, verification gates, anti-rationalization tables (common excuses paired with rebuttals), and red-flag indicators. There are also expert personas (code-reviewer, test-engineer, security-auditor) and reference checklists.
The distinguishing idea
The README is explicit that the skills encode practices from Google’s engineering culture, citing ideas like Hyrum’s Law, the Beyonce Rule for testing, Chesterton’s Fence for simplification, and trunk-based development. That is the differentiator worth dwelling on: where many skill collections are personal preference, this one claims a documented engineering tradition as its source. Whether that resonates depends on whether you want your agent nudged toward big-company rigor (specs, tests, reviews, security gates) or toward speed.
Install
# Claude Code
/plugin marketplace add addyosmani/agent-skills
/plugin install agent-skills@addy-agent-skillsBeyond Claude Code it supports Cursor (copy SKILL.md to .cursor/rules/), Gemini CLI, Windsurf, OpenCode, Copilot, and Kiro, plus a generic markdown system-prompt path. Seven top-level commands map the lifecycle: /spec, /plan, /build, /test, /review, /code-simplify, /ship. A /build auto mode generates a plan and implements tasks one slice at a time, pausing on risk.
When it fits, and when it does not
It fits teams that want their agent to behave like a disciplined senior engineer: write the spec, prove it with tests, review for quality and security before shipping. It fits less well for throwaway scripts or rapid prototyping, where the gates add friction you do not want. It is opinionated toward reliability over speed, which is the right trade for production code and the wrong one for a weekend hack.
How it compares
| Project | Angle | Stars (2026-06) |
|---|---|---|
| addyosmani/agent-skills | Google-practice quality gates | ~52k |
| obra/superpowers | Skills framework + methodology | ~220k+ |
| anthropics/skills | First-party agent skills | first-party baseline |
| mattpocock/skills | One engineer’s .claude directory | personal, curated |
superpowers is the broader methodology framework; anthropics/skills is the official baseline; mattpocock/skills is a personal directory shared. agent-skills’ niche is the explicit “this is how Google engineers work” framing plus the verification gates and anti-rationalization tables, which push back when the agent tries to cut corners.
Gotchas from the issue tracker
The project is young (created early 2026) and the open issues are mostly integration and routing rather than core defects:
- A plugin path containing spaces could break the SessionStart hook (#214).
- There is a documented routing ambiguity between the code-reviewer persona and the code-review-and-quality skill (#173), so the meta-skill’s routing is still being tightened.
- Several issues request first-class support for additional hosts (for example Antigravity, #226, #177).
The takeaway: the skills themselves are solid, and the friction is in install paths and in routing among 23 overlapping skills. The meta-skill exists precisely to manage that, and it is still maturing.
FAQ
Is addyosmani/agent-skills free? Yes. agent-skills is MIT-licensed and open-source, with no paid tier.
How many skills does agent-skills have? 23: 22 lifecycle skills spanning define, plan, build, verify, review, and ship, plus one meta-skill that routes between them.
Does agent-skills work with Cursor and Gemini? Yes. Beyond Claude Code it supports Cursor, Gemini CLI, Windsurf, OpenCode, Copilot, and Kiro, plus a generic markdown system-prompt path.
How do I install agent-skills in Claude Code? Run /plugin marketplace add addyosmani/agent-skills then /plugin install agent-skills@addy-agent-skills.
What makes addyosmani/agent-skills different from other skill collections? It explicitly encodes practices from Google’s engineering culture (Hyrum’s Law, the Beyonce Rule, trunk-based development) and ships verification gates plus anti-rationalization tables that push back when the agent tries to skip a test or cut a corner. The bias is reliability over speed, which suits production code more than quick prototypes. The author’s background as a Google engineering director and web-performance authority is also a large part of why the repo gathered tens of thousands of stars so quickly, so weigh the content on its merits rather than the name alone.
Related reading
Compare the broader framework in obra/superpowers, the first-party baseline in anthropics/skills, and a personal collection in mattpocock/skills. For a far broader role library, see msitarzewski/agency-agents.