What it is
Agent Skills is Addy Osmani's collection of Markdown skills that turn senior-engineer habits into steps an agent must follow. Each skill has a process, checkpoints, a table of common excuses with rebuttals, red flags, and the evidence required before the agent can call the work done. Many rules come from Google's engineering culture, such as Hyrum's Law and Chesterton's Fence.
The pack has 25 skills (24 lifecycle skills plus a using-agent-skills router) grouped by stage (define, plan, build, verify, review, ship), 9 slash commands (/spec, /plan, /build, /test, /constraints, /review, /webperf, /code-simplify and /ship), 4 reviewer personas and 7 shared checklists. It installs as a Claude Code or Codex plugin, into Gemini CLI and Antigravity, or into 70+ agents through the skills CLI. Release 0.6.13 shipped on 13 October 2026.
Our review
Agent Skills is the pack for an agent whose problem is discipline rather than knowledge: one that writes code before there is a spec, declares victory without running the tests and ships a 600-line diff. Every skill ends in evidence requirements and carries a table of the excuses agents use to skip steps, each with a rebuttal. At about 109,800 GitHub stars it is among the most-starred skill packs, though Jesse Vincent's Superpowers, at about 304,000, is larger.
Its distinctive ground is web work. browser-testing-with-devtools uses Chrome DevTools MCP for DOM, console, network and performance data, /webperf runs a Core Web Vitals audit through a dedicated persona, and the front-end skill sets WCAG 2.1 AA as the bar; reviewers at Productize found this the one area its trending rivals skipped. /build auto asks for plan approval once, then implements each task with tests and its own commit, stopping on failure. Skill routing is tested in CI against 25 eval case files.
The rough edges are documented, which counts in its favour. In Claude Code the built-in /review wins, so you type /agent-skills:review. A single-skill npx install drops the shared references/ checklists the skills point to (issue #361). The Claude Code marketplace clones over SSH and fails without a GitHub key unless you give it the HTTPS URL. The hooks are opt-in and documented in their own files: a session-start router the plugin deliberately leaves unwired, and a WebFetch cache that revalidates with the origin server on every reuse.
Breadth has a context cost. Osmani advises starting with the four or five skills closest to your problems, and running the pack beside Superpowers hands the agent two competing rulebooks for planning, TDD and review. Matt Pocock's skills and GitHub's Spec Kit each own one slice of the workflow; this pack tries to own all of it, which is the reason to adopt it and also the reason to adopt only part.
Where it earns its place
A front-end lead on a 15-person e-commerce team using Claude Code
Stop agent-written pages regressing Core Web Vitals and accessibility
Install the plugin and require /test with browser-testing-with-devtools and a /webperf audit before merge; both read real console, network and performance data through Chrome DevTools MCP. The pack is free; the cost is longer agent sessions on the existing Claude plan.
A solo developer starting a greenfield SaaS app
Get from idea to a reviewed first release without skipping the spec
Run /spec for a PRD, /plan to split it into tasks, then /build auto, which asks for plan approval once and implements each task with tests and a separate commit, stopping on failure. The adoption guide recommends the full lifecycle for new projects and a gradual rollout for established ones.
An engineering manager rolling out agents to 40 developers on Codex and Cursor
Give every agent the same review and security standards
Install the shared skills/ core per tool (Codex plugin, .cursor/skills) and use /constraints to write CONSTRAINTS.md with the team's thresholds once. MIT licence and no per-seat fee; budget time to trim the pack, since all 25 skill descriptions compete for context.
Before you commit
- Setup
- `npx skills add addyosmani/agent-skills` installs all 25 skills into most agents. In Claude Code, `/plugin marketplace add addyosmani/agent-skills` then `/plugin install agent-skills@addy-agent-skills`, using the HTTPS URL without an SSH key. Codex needs CLI v0.122+.
- Integrations
- Claude Code, Codex, Gemini CLI, Antigravity, Copilot CLI, Cursor, Windsurf, OpenCode, Kiro and Command Code; Chrome DevTools MCP for browser testing. The lifecycle slash commands exist only where the host supports them; in Copilot you invoke skills by name.
- Data and privacy
- Plain Markdown plus shell-script hooks that run inside your agent; the pack itself sends nothing anywhere. The optional sdd-cache hook stores fetched documentation on disk and revalidates it with the server before reuse. MIT licence.
- Cost at scale
- Free. The cost is tokens: full-lifecycle runs with specs, tests and reviews are longer sessions than an unguided agent's, and installing every skill loads more descriptions into context. The project publishes no token figures.
- Limits
- Command names can clash with host built-ins; single-skill installs lose references/; some Antigravity releases hide the wrapper commands; overlaps with Superpowers on planning, TDD and review.
What's good
- Every skill ends in evidence requirements
- Chrome DevTools browser testing and /webperf audits
- Installs into Claude Code, Codex, Gemini and 70+ agents
- Skill routing tested in CI with 25 eval cases
What's not
- /review clashes with Claude Code's built-in
- Single-skill installs lose the shared checklists
- Marketplace install fails without a GitHub SSH key
- Conflicts with Superpowers if both are installed
Best for
Web developers who want one structured skill pack from spec to launch.
Alternatives
Agent Skills (Addy Osmani): common questions
- Is Agent Skills (Addy Osmani) worth it?
- Addy Osmani's pack turns senior-engineer habits into gates an agent cannot talk its way past, and it is the rare skill set that takes the browser seriously. We rate it 4.3 out of 5.
- How much does Agent Skills (Addy Osmani) cost?
- Free and open source (MIT). You pay for your agent's model usage.
- What are the downsides of Agent Skills (Addy Osmani)?
- /review clashes with Claude Code's built-in; single-skill installs lose the shared checklists; marketplace install fails without a GitHub SSH key; conflicts with Superpowers if both are installed.
- Who is Agent Skills (Addy Osmani) best for?
- Web developers who want one structured skill pack from spec to launch.
- How much does Agent Skills (Addy Osmani) cost for a team or in production?
- Free. The cost is tokens: full-lifecycle runs with specs, tests and reviews are longer sessions than an unguided agent's, and installing every skill loads more descriptions into context. The project publishes no token figures.
- Is Agent Skills (Addy Osmani) safe for company data?
- Plain Markdown plus shell-script hooks that run inside your agent; the pack itself sends nothing anywhere. The optional sdd-cache hook stores fetched documentation on disk and revalidates it with the server before reuse. MIT licence.
- What are the best Agent Skills (Addy Osmani) alternatives?
- The closest alternatives we have reviewed are Superpowers, Matt Pocock's Skills and Spec Kit.
Answers checked Oct 23, 2026.
Make Agent Skills (Addy Osmani)? Show this rating on your site
