Agent Skills: Making Claude Coding Agents Follow Senior‑Engineer Workflows
Add the Agent Skills slash commands to your Claude Code environment to enforce spec, test, review workflows.
Install the Agent Skills plugin via the Claude Code marketplace and enable the /spec, /plan, /build, /test, /review, /ship, and /code‑simplify commands for your projects.
Summary
The Agent Skills project, now over 26 000 GitHub stars, adds a “senior‑engineer” scaffolding to Claude‑based AI coding agents so they no longer just spit out code.
A skill is a markdown file with frontmatter that the agent injects into its context, turning a simple prompt into a step‑by‑step workflow that ends with concrete evidence such as a passing test or a signed review.
The library ships 20 distinct skills that map to the six SDLC phases—Define (/spec), Plan (/plan), Build (/build), Verify (/test), Review (/review), Ship (/ship)—and exposes seven slash commands that trigger the appropriate skill.
Each skill contains an anti‑rationalization table that pre‑writes rebuttals to common excuses, ensuring the agent writes a failing test first, refactors only what is asked, and never skips verification.
The design also enforces progressive disclosure: only the skill relevant to the current phase is loaded into the agent’s context, keeping token usage low while still providing full workflow coverage.
Scope discipline is baked into the meta‑skill so the agent will not touch unrelated files or remove TODOs without a clear plan, mirroring Google’s code‑review norms.
The project explicitly incorporates Google engineering practices such as Hyrum’s Law, the test pyramid, the Beyoncé rule, and trunk‑based development, all encoded as workflow checkpoints.
By installing the Agent Skills plugin via the Claude Code marketplace, teams can automatically route feature requests through these structured steps and reduce the risk of shipping fragile code.
Key changes
- 20 skills covering 6 SDLC phases with 7 slash commands (/spec, /plan, /build, /test, /review, /ship, /code-simplify)
- Anti‑rationalization tables embedded in each skill to counter common excuses
- Progressive disclosure: only the active skill is loaded into context to keep token usage low
- Scope discipline meta‑skill prevents the agent from touching unrelated code
- Verification checkpoints require passing tests, clean build, runtime trace, and reviewer sign‑off
- Google engineering practices (Hyrum’s Law, test pyramid, Beyoncé rule, trunk‑based dev) are encoded as workflow steps
- The library is distributed as a GitHub repo with 26 000 stars and can be installed via Claude Code marketplace
- The slash commands trigger the appropriate skill automatically when the agent is in the corresponding phase