agent-skills (Addy Osmani) — Production Engineering Skills for AI Coding Agents

Source: raw/gh-star-addyosmani-agent-skills.md + ai-research/addyosmani-agent-skills-readme-2026-06-13.md (README, 2026-06-13) — independently corroborated by Matthew Berman’s OSS-tools video (raw/You_NEED_to_try_these_open-source_AI_projects_RIGHT_NOW.md) and an X bookmark (raw/x-bookmarks-recent-digest-2026-06-13.md).

addyosmani/agent-skills is a production-grade, multi-IDE collection of engineering skills for AI coding agents, authored by Addy Osmani (Google Chrome engineering lead, well-known web-performance author). At ~58,000 GitHub stars (MIT) it is one of the highest-signal skill collections in the ecosystem. It packages the workflows, quality gates, and best practices senior engineers use into 7 lifecycle slash commands and 16 skills that activate automatically by phase — and ships first-class setup paths for Claude Code, Cursor, Antigravity, Gemini CLI, Windsurf, OpenCode, GitHub Copilot, Kiro, and Codex.

Key Takeaways

• A spec-driven lifecycle, not a grab-bag. The 7 commands map to DEFINE → PLAN → BUILD → VERIFY → REVIEW → SHIP: /spec (spec before code), /plan (small atomic tasks), /build (one vertical slice at a time), /test (tests are proof), /review (improve code health), /code-simplify (clarity over cleverness), /ship (faster is safer). Skills also auto-activate by context — designing an API triggers api-and-interface-design, building UI triggers frontend-ui-engineering.
• /build auto is the autonomous mode that keeps the verifier. It generates the plan and implements every task in a single approved pass — removing the human between tasks, not the verification. Each task stays test-driven and is committed individually; it pauses on failures or risky steps. This is the maker/verifier discipline the verification frontier thesis argues for, packaged as a command.
• 16 skills spanning the full lifecycle. Notables: interview-me (one-question-at-a-time interview to ~95% confidence before building), spec-driven-development (PRD before code), test-driven-development (Red-Green-Refactor, 80/15/5 test pyramid, DAMP-over-DRY), doubt-driven-development (adversarial fresh-context review: CLAIM → EXTRACT → DOUBT → RECONCILE → STOP, with optional cross-model escalation), source-driven-development (ground every framework decision in official docs and cite it), context-engineering, code-review-and-quality (five-axis review, ~100-line change sizing, Nit/Optional/FYI severities), and code-simplification (Chesterton’s Fence, Rule of 500).
• Framework-agnostic by design. Skills are plain SKILL.md markdown, so they run with any agent that accepts instruction files. The repo documents concrete install paths for 9 surfaces — in Claude Code it installs as a plugin via the marketplace (/plugin marketplace add addyosmani/agent-skills → /plugin install agent-skills@addy-agent-skills).
• Where it sits among skill collections. Broader and more lifecycle-complete than single-author bundles like Matt Pocock’s skills, and a curated counterpart to the open registry skills.sh. It overlaps conceptually with Anthropic’s own skills guidance but is opinionated about the whole spec→ship workflow, not individual capabilities.

Try It

• Install in Claude Code and run the spine. /plugin marketplace add addyosmani/agent-skills then /plugin install agent-skills@addy-agent-skills. On your next feature, run /spec → /plan → /build → /test → /review → /ship once end-to-end to feel the lifecycle before customizing.
• Use interview-me on an underspecified task. Before letting an agent build from a vague ask, invoke the interview skill — it extracts what you actually want one question at a time. This is the cheapest reliability win in the bundle.
• Cherry-pick into an existing setup. Skills are plain markdown — copy doubt-driven-development or code-review-and-quality into .cursor/rules/ or your AGENTS.md without adopting the whole plugin.
• Try /build auto on a low-stakes change first. Confirm it still pauses on failures and commits per-task before trusting it on anything production-touching — the same self-guard discipline Fable 5’s failure-mode taxonomy warrants.

Related

• skills.sh — Open Agent Skills Directory — the open registry; agent-skills is a curated, lifecycle-complete collection rather than a directory.
• Matt Pocock’s skills — a narrower single-author skill bundle; useful contrast on scope.
• Building Agents With Skills — Anthropic’s official skills guidance that collections like this build on.
• Skills Ecosystem — the open-standard context agent-skills plugs into across 9 IDEs/agents.
• Headroom — the sibling tool surfaced in the same Matthew Berman OSS-projects video.
• The 2026 Claude Code AIOS Pattern — the operator pattern whose “skill as the capability primitive” dimension this collection instantiates at production grade.
• Five OSS Tools That Fix Claude Code’s Blind Spots — adjacent operator-workflow toolkit.