skill-reviewer
Based on Anthropic's Claude Code Skills architecture. A reviewer for the skills you write on that foundation.
You built a skill. You want feedback. You ask a friend. Your friend says: "Looks great! Maybe consider adding tests?" That's not a review. That's a hug.
skill-reviewer reads every file, cites specific lines, and separates strengths from weaknesses with severity-rated finding cards. It refuses to hedge: if something is fragile, it'll say so; if something is unusually well-done, it'll say that too.
~4 min read · scan the TL;DR if you only have 30 seconds
TL;DR
- What: A Claude Code skill that reviews other Claude Code skills with file:line citations, severity-rated finding cards, and an optional second-opinion challenger pass.
- Why: Polite-by-default feedback ships skills with quiet flaws. skill-reviewer gives you the candid review you'd want before publishing.
- Install: Two
/plugin commands in Claude Code (below); then /skill-reviewer is available in any session.
- Try first:
/skill-reviewer summary /path/to/some-skill — 5-minute sanity check (TL;DR + top 5 strengths + top 5 weaknesses). Lighter than a full review.
- Example output: a real review of the
unforget skill.
- Maturity: v0.3.1; five releases in the v0.x line; self-reviews after every meaningful edit.
Newer to Claude Code?
A skill is a markdown file Claude Code knows how to run. When you type /skill-reviewer review <path>, Claude follows the instructions in this skill, reads the target skill's files, and writes you a structured report. You don't have to memorize anything — the skill tells Claude what to do, you read the report.
A real fragment
This is the top of a real review (skill-reviewer reviewing the unforget skill). It's the kind of output you'll get on your own work:
## TL;DR
`unforget` is a Claude Code skill that consolidates deferred work into one
structured UNFORGET.md per project. It's more disciplined than most skills
I've reviewed: thin SKILL.md → fat reference files, deterministic Python
helpers with prose fallbacks, backups before destructive operations, and the
verify-still-open recipe is a genuinely original idea worth generalizing.
The standout strength is structural; the standout weakness is that the
canonical example never demonstrates the project's signature feature.
Weakness clusters:
1. Tooling and parseability gaps — Compact preset has no canonical regex
2. Discoverability — natural-language trigger phrases buried in YAML
3. Test coverage — Surface 6 and prune_backups.py are explicitly untested
4. Duplicate sources of truth — closure pointer format lives in 3 places
5. Under-specified fallbacks — "fail-soft" and Python-missing triggers vague
Notice what's there: file paths, named gaps, specific patterns. Notice what's not: "great work overall," "minor improvements possible," "consider thinking about." The full report (preserved at skills/skill-reviewer/examples/sample-report-unforget.md) continues with the strengths to keep, individual finding cards (each with Why / Fix / citation), and a free-form Patterns section synthesizing what crossed cards.
Install
Two commands in any Claude Code session, one at a time. Paste both at once and the second gets eaten as arguments to the first. Ask me how I know:
/plugin marketplace add Terryc21/skill-reviewer
Wait for "Successfully added marketplace," then:
/plugin install skill-reviewer@skill-reviewer
Verify with /skill-reviewer --version. You should see three lines: version, install path, and the list of available lenses.
How invocation works (autocomplete caveat)
The plugin registers /skill-reviewer as a single skill activation. Claude parses everything after it as a subcommand string — review <path>, summary <path>, lenses, detect <path>, --version. The subcommands listed throughout this README are recognized argument shapes, not separately-registered slash commands. There is no per-subcommand autocomplete; typing /skill-reviewer and pressing Tab won't enumerate review / summary / lenses for you. (Tab will offer /skill-reviewer itself once the plugin is installed.)
