terryc21/skill-reviewer

skill-reviewer is itself a Claude Code skill, and its job is to review other Claude Code skills. A skill, in this setting, is a folder with a markdown file and supporting files that tells Claude how to do a particular task. The README opens with a complaint: when you ask a friend to look at a skill you have written, you usually get a polite answer like, looks great, maybe add some tests. The author calls that a hug, not a review. skill-reviewer is meant to read every file in your skill, cite specific lines, separate strengths from weaknesses, and rate findings by severity instead of hedging. Using it inside Claude Code is straightforward. There are two install commands, both for the /plugin marketplace, that have to be run one at a time. After that, /skill-reviewer is available in any session. The simplest invocation is /skill-reviewer review followed by the path to a skill directory. After two to five minutes, you get a markdown report with a short TL;DR, five to ten strengths to keep, a sorted stack of finding cards (each one with a Why, a Fix, a file and line citation, an effort estimate, and a quick-win tag), and a Patterns section that names themes that cross several findings. For a lighter pass there is also /skill-reviewer summary, which takes about five minutes and returns the headline verdict with three to five strengths and up to five finding cards. The README lists a few other ways to scope a run. You can pass --lens=safety, or another lens, to focus the review on one concern. You can pass --second-opinion to add a challenger pass that pushes back on the first verdict, which the author recommends before publishing. There is also a /skill-reviewer detect command that just classifies the shape of a skill and a /skill-reviewer lenses command that lists the available lenses. A manual install path exists for users who cannot use the plugin marketplace, in which case the invocation becomes /skill skill-reviewer followed by the same arguments. The README is honest about a few details. There is no per-subcommand autocomplete, so tab-complete after /skill-reviewer will not list review, summary, and so on. Every review is a fresh read of the files at the given path; the skill does not auto-load older reports, although prior reports are kept in a .agents/research/ folder for manual reference. The author also notes that the project self-reviews after every meaningful edit, and is at version 0.3.1 with five releases in the 0.x line. The project is positioned for moments before publishing a skill, when auditing someone else's skill, before merging a large pull request to a skill, or when you suspect your own internal review has been either too kind or too harsh.