Agent Skills (Addy Osmani)
An independent collection of production-grade engineering Agent Skills that walk AI coding agents through a structured software development lifecycle. Built by Addy Osmani (Google Chrome eng lead, author of Learning JavaScript Design Patterns) with Federico Bartoli and Joan León. Not affiliated with Anthropic’s skills repo — it is a methodology stack, sibling to Superpowers and gstack.
Encodes Google engineering culture into skills: Hyrum’s Law, test pyramids, trunk-based development, anti-rationalization tables that tell the agent when it’s allowed to skip a step.
Links
Description
GitHub: addyosmani/agent-skills — MIT, ~77k stars / ~8k forks. Cross-harness: same SKILL.md runs on 70+ agents (Claude Code, Cursor, Codex, Gemini CLI, …).
Download or use
Install like any skill collection — copy into the harness skills path (.claude/skills/ for Claude Code; see Awesome Agent Skills cheat-sheet for other harnesses), or clone the repo. Skills auto-trigger by description; slash commands drive each phase explicitly.
Reasoning for
Most skill collections are ad-hoc single-purpose helpers. This one is opinionated end-to-end SDLC — 24 skills mapped onto lifecycle phases so the agent follows real engineering process instead of vibe-coding to done.
The 24 skills, by phase
| Phase | Skills | What it enforces |
|---|---|---|
| Define | 3 | Spec/requirements before code |
| Plan | 1 | Explicit plan from the spec |
| Build | 7 | Implementation with guardrails |
| Verify | 2 | Tests, checks |
| Review | 4 | Code review, simplification |
| Ship | 7 | Release, observability, cleanup |
Plus 1 meta-skill for navigation.
Slash commands (phase entry points)
/spec → /plan → /build → /test → /review → /webperf → /code-simplify → /ship
/build auto = optional autonomous mode — end-to-end task execution with per-step verification.
Specialist personas & checklists
- 4 agent personas — code reviewer, test engineer, security auditor, web-performance auditor
- 7 reference checklists — testing, security, performance, accessibility, observability (loaded on-demand = progressive disclosure)
Alternatives considered
- Superpowers — Jesse Vincent’s (obra) TDD-first methodology stack (brainstorm→worktree→plan→subagent→TDD→review→finish); more test-driven, multi-host
- gstack — Garry Tan’s (YC) virtual-team stack (CEO/Designer/EngMgr/QA/SRE), sprint-driven
- Karpathy Skills — a single CLAUDE.md instead of a multi-skill lifecycle
- Awesome Agent Skills — the curated superset to discover these and more
Resources
- GitHub: addyosmani/agent-skills
- Agent Skills — the underlying mechanism and open standard
- Superpowers / gstack — peer SDLC/virtual-team skill stacks
- Karpathy Skills — single-file counter-approach
- Awesome Agent Skills — curated 1000+ skills list (includes this)
- Harness Engineering — configuring the harness these skills run in
- Claude Code — primary harness
Template: tool