Name
agent-skills
Agent skills that work in Claude Code, Codex and other AI coding agents from one copy, for reviewing changes, writing pull requests, Bash scripts and maintaining skills.
Synopsis
- status
- available
- role
- Author and maintainer
- period
- 2026–
- stack
- Markdown, JavaScript, Node.js, Bash
- source
- github.com/winmetta/agent-skills
Story
Who needed help
I use more than one AI coding agent, and so do the volunteers I work with. Each agent had its own way of loading instructions, so the same review checklist or PR template got copied, edited in one place and forgotten in the other.
What I built
A catalog of skills in the open Agent Skills format, written so one copy works in every agent:
- local-change-review reviews uncommitted work before a commit, checks each finding again in a separate pass, and fixes what it finds until a review comes back clean.
- local-pr-review reviews a pull request locally and posts the findings as a GitHub App bot, so the author has to resolve each thread.
- pr-description writes PR titles and descriptions with the same sections every time, reporting only checks that actually ran.
- bash-scripting writes scripts as small functions and validates them with ShellCheck and shfmt.
- skill-maintainer creates and tests skills, including a trigger test against the real Claude Code CLI.
The tradeoffs that mattered
Evidence over confidence. The review skills only report a finding with a concrete failing case, and a second, independent pass has to agree. Fewer findings, but people stop ignoring them.
No dependencies. The helper scripts need only Node.js, so installing a skill never pulls in a package tree.
Where it stands
Every repository I maintain uses these skills before each commit and on each pull request. Install one with npx skills@latest add winmetta/agent-skills --skill <name>.