Agent Skills: AI Agent Skills Collection for Coding
addyosmani/agent-skillsAgent Skills is an open-source AI agent skills collection from Addy Osmani: 24 structured workflows that push AI coding agents through the spec-plan-build-test-review-ship discipline a senior engineer would follow. Instead of one-off prompts, each skill is a step-by-step process with verification gates, so the agent writes a PRD before code and proves tests pass before shipping.
Reach for Agent Skills if you want your coding agent to stop taking the shortest path — skipping specs, tests, and reviews — and instead follow a repeatable spec-to-ship workflow across whatever agent you use. It's content you wire into your agent, not an app that runs, so the payoff depends on the agent honoring the skills. Skip it if you already maintain your own hardened prompt/skill setup, or you want something that does the work rather than governing how it's done.
The problem it solves
AI coding agents default to the fastest route to 'done,' which usually means no spec, thin tests, and no security or performance pass. That's fine for a throwaway prototype and risky for anything that ships. Without a shared process, teams end up re-explaining the same engineering discipline to the agent on every task instead of encoding it once.
What is it?
Agent Skills is an open-source AI agent skills collection: 24 Markdown-defined workflows (23 lifecycle skills plus one meta-skill) packaged so AI coding agents follow senior-engineer practices consistently. It's MIT-licensed, curated by Addy Osmani, and — per the README — installs into 70+ agents including Claude Code, Cursor, Codex, Copilot, and Cline via the open skills CLI (`npx skills add addyosmani/agent-skills`). Each skill is a process with steps, an anti-rationalization table, and a verification gate rather than loose prose.
Why it's getting attention
At 77k+ stars it's one of the more visible agent-skills packs, and it rides two currents: the Vercel skills CLI that makes skill packs installable across 70+ agents, and the broader move toward giving agents reusable workflows instead of ad-hoc prompts. The README ties the skills to Google's engineering practices — Software Engineering at Google and the public eng-practices guide — which gives the discipline a recognizable pedigree.
Key features
- ✓24 skills covering the full lifecycle — spec, plan, build, test, review, ship — plus a meta-skill that routes work to the right one
- ✓8 slash commands (/spec, /plan, /build, /test, /review, /webperf, /code-simplify, /ship) that activate the relevant skills automatically
- ✓A `/build auto` mode that generates the plan and implements every task in one approved pass, still test-driven and committed per task
- ✓4 specialist review personas: code reviewer, test engineer, security auditor, and web-performance auditor
- ✓7 reference checklists — definition-of-done, testing, security, performance, accessibility, observability, and orchestration
- ✓Cross-agent install through the open skills CLI (Claude Code, Cursor, Codex, Copilot, Cline, and 70+ others), plus native plugin paths for several tools
- ✓Anti-rationalization tables and a verification gate in each skill, so 'seems right' isn't accepted as done
Best use cases
- •Forcing a spec or PRD before the agent writes any code on a new feature
- •Wiring red-green-refactor TDD into an agent's default build loop
- •Running a five-axis code review or a security/performance audit before merge
- •Standardizing how a whole team's agents behave across Claude Code, Cursor, and Codex
- •Cherry-picking a single skill (e.g. code-review-and-quality) into an existing setup
How to install / try
Fastest path is the open skills CLI: `npx skills add addyosmani/agent-skills` installs all 24 skills into 70+ agents; add `--list` to browse first or `--skill <name>` to grab one. For Claude Code natively, the README gives `/plugin marketplace add addyosmani/agent-skills` then `/plugin install agent-skills@addy-agent-skills`. Native setup guides exist for Cursor, Codex, Antigravity, Gemini CLI, Windsurf, OpenCode, and Copilot.
How to use
Once installed, skills activate automatically from what you're doing — designing an API triggers the api-and-interface-design skill, building UI triggers frontend-ui-engineering — or you invoke a slash command like `/spec` or `/review`. You can also reference any skill by name directly. Each skill is plain Markdown, so it works with any agent that accepts instruction files.
Strengths
- ✓Opinionated, process-driven workflows with steps, checkpoints, and exit criteria — not generic 'be a good engineer' prompts
- ✓One install covers 70+ agents through the skills CLI, so you're not locked to a single tool
- ✓Plain Markdown and MIT-licensed, so you can read, fork, and edit every skill
- ✓Verification gates and anti-rationalization tables push back on the agent quietly skipping tests or reviews
- ✓Modular — install all 24 or cherry-pick the individual skills you want
Limitations & risks
- △This is a collection of instruction files, not a runnable tool — you wire it into your own agent, and its value depends on that agent actually following the skills
- △No enforcement outside the model: an agent can still ignore a skill's verification gate; the tables discourage it but nothing hard-blocks it
- △The skills encode a specific, Google-flavored engineering opinion that may not match how your team already works
- △Cross-agent behavior varies — a skill that assumes Chrome DevTools MCP or CI integration only helps if you have that wired up
Alternatives
Who should try it — and who should skip
Teams and individuals who use AI coding agents daily and want them to follow consistent engineering discipline — spec first, tests as proof, review before merge — without re-prompting it every task. It's most useful if you run more than one agent (say Claude Code plus Cursor or Codex) and want the same workflow across all of them. If you don't want an opinionated process layer, or you already maintain your own skill/prompt library, you can skip it.
Frequently asked questions
Agent Skills is an open-source, MIT-licensed collection of 24 engineering workflows for AI coding agents. Each skill is a Markdown-defined process — spec, test, review, ship — that an agent follows instead of taking the shortest path to done.
Run `npx skills add addyosmani/agent-skills` to install all 24 skills into 70+ agents via the open skills CLI, or use the `/plugin marketplace add addyosmani/agent-skills` flow for Claude Code natively.
Yes. The README says the skills CLI installs into 70+ agents including Cursor, Codex, Copilot, and Cline, with native setup guides for several; the skills are plain Markdown that any instruction-following agent can use.
Yes — it's open source under the MIT license, so you can use, fork, and modify every skill in the pack.
No. It's a collection of instruction files you wire into your own agent; it doesn't run on its own, and its value depends on your agent honoring the skills' steps and verification gates.