liu-penpen/skill-review

This repository is a tool, itself packaged as an "Agent Skill," for reviewing the quality of other Agent Skills, which are structured instruction packages used by AI coding assistants such as Cursor or Claude. The author's stated goal is that the review tool must follow the same quality rules it checks for in other skills, so it does not end up being low effort output used to judge other low effort output. The core of the project is a set of 10 checkable rules, called a rubric, covering things like whether a skill loads its instructions gradually instead of all at once, whether its description uses the right keywords so it gets triggered correctly, whether it includes a clear workflow checklist, whether it has confirmation steps at the right points, whether scripts are wrapped properly, how arguments are handled, how supporting reference files are organized, whether there is a pre delivery self check, and whether the skill offers both a command line and an in chat interface. A separate rule looks specifically for signs of low quality, filler generated content. Alongside the rubric, the package ships a small script, run with Node.js, that checks these rules automatically without needing any external dependencies, and produces a report with findings graded from P0, the most severe, down to P3. The tool can review an existing skill and produce that graded report, generate a new skill folder from a template that already passes the lint checks, focus on a specific subset of rules, or run a faster check that skips the more subjective review steps. Installation is a single command using a tool called npx skills, and once installed the skill can be triggered inside a chat session with a slash command, natural language requests like asking to review or audit a skill at a given path, or run directly from the command line for use in continuous integration pipelines or git hooks. The README documents each of these three ways to use it, along with example commands for each.