agent-skills: 25 engineering disciplines as agent skills

Coding agents take the shortest path by default: skip the spec, skip the tests, skip the security review, and forget the rules you set last turn. agent-skills, open-sourced by addyosmani (Google Chrome team) in February 2026 under MIT (about 106k stars and 11k forks as of 2026-10-13), attacks that habit directly. It does not swap the model; it writes the workflows, quality gates and best practices senior engineers use into 25 SKILL.md files, so agents like Claude Code, Codex and Cursor follow the same discipline at every phase of development.

Core features

• 25 skills across the lifecycle: 24 lifecycle skills plus the using-agent-skills meta-skill. The define phase ships interview-me (one question at a time until ~95% confidence) and spec-driven-development (PRD before code); planning uses planning-and-task-breakdown; building proceeds in vertical slices; verification runs on test-driven-development; code-review-and-quality and security-auditor gate the merge.
• Nine slash commands as entry points: /spec defines, /plan breaks down, /build implements, /test proves, /constraints sets the quality bar, /review checks health, /webperf audits performance, /code-simplify clarifies, /ship releases — each activates the right skills automatically. /build auto runs every task after one approved plan, pausing on failures or risky steps.
• Anti-rationalization tables are the signature: every skill carries a “Common Rationalizations” table that lists the excuses agents use to skip steps, each with a written rebuttal. In spec-driven-development, “this is simple, I don’t need a spec” is answered with “simple tasks don’t need long specs, but they still need acceptance criteria.”
• Verification is non-negotiable: every skill ends with an evidence checklist — tests passing, build output, runtime data — and “seems right” is never sufficient. spec-driven-development even makes “stop your turn immediately after writing the spec” a hard rule; planning starts only after the human approves in a later turn.
• Four prebuilt agent personas: code-reviewer (five-axis review), test-engineer (test strategy), security-auditor (threat modeling) and web-performance-auditor (Core Web Vitals), invocable per task.
• One command installs into 70+ agents: the open skills CLI covers Claude Code, Cursor, Codex, Copilot and Cline; a Claude Code marketplace plugin, a Cursor rules directory and an Antigravity CLI plugin are also supported.

Typical use cases

• A team whose agent-written code runs but nobody dares review it: write the gates into CONSTRAINTS.md with /constraints and enforce them before every commit.
• Underspecified feature requests: let interview-me extract what the user actually wants, then move to /spec.
• Incremental adoption on an existing codebase: follow the Adoption Guide’s verification-first route instead of a big-bang rollout.
• Architectural decisions nobody writes down: documentation-and-adrs forces the “why” into an ADR, so the next agent touching that code reads the decision instead of reinventing it.

Quick start

npx skills add addyosmani/agent-skills            # install all 25 skills
npx skills add addyosmani/agent-skills --list     # browse before installing

Single skills install too, e.g. npx skills add addyosmani/agent-skills --skill test-driven-development. In Claude Code, typing /spec shows the workflow in motion. Skills also activate on their own based on the work in front of the agent — designing an API pulls in api-and-interface-design, building UI pulls in frontend-ui-engineering — so you do not have to memorize names. Every integration reads the same skills/ directory; the CLI, plugin and rules files are only entry points.

Summary

Aimed at developers already shipping production code with coding agents and tired of the skipped steps; occasional scripters may find the process heavy. How many of those 106k stars are actually in use is unknowable, but the problem it treats is real — install it if your team has that problem. Three catches up front: installing a single skill drops the repo-level references/ checklists (issue #361); the Claude Code marketplace clones over SSH, so a missing GitHub key surfaces as Permission denied; and some Antigravity CLI releases report legacy commands as converted while their wrappers stay undiscoverable. For more Claude Code ecosystem resources, everything-claude-code is a maintained list; animate shows a different way to constrain an agent — writing “confirm before rendering” into the skill itself.