Agent Skills

Addy Osmani's Agent Skills, which has surpassed 26K stars, is Markdown-based workflow scaffolding that prevents AI coding agents from skipping senior engineering procedures such as writing specs, writing tests first, producing reviewable PRs, and reviewing trust boundaries.

Skills are less like reference documents and more like executable procedures with step ordering, checkpoint evidence, and exit criteria. The repository bundles 20 skills into 6 stages—Define / Plan / Build / Verify / Review / Ship—along with 7 slash commands: /spec, /plan, /build, /test, /review, /ship, and /code-simplify.

The core principles are as follows.

• Process over prose: include executable workflows rather than essays.
• anti-rationalization tables: preemptively rebut excuses like "this task is too simple."
• Verification as exit criteria: use not just passing tests but runtime behavior, logs, and review approval as evidence.
• progressive disclosure: load only the skills that are needed to reduce context pollution.
• scope discipline: touch only what was requested, not adjacent systems.

The article connects this structure to public engineering practices such as Google's design docs, code review, trunk-based development, feature flags, and the DAMP testing philosophy. It notes that the skills can be ported as Markdown files into Claude Code, Cursor, Gemini CLI, Codex, Aider, Windsurf, OpenCode, and others, and that even without installing them, you can bring the anti-rationalization rules, verification criteria, and progressive disclosure into your AGENTS.md and team runbooks.