Agent Skill Reviewer and Creator

The problem

Agent Skills are often governed by prose conventions, so issues like weak description triggers, missing confirmation checkpoints, nondeterministic scripts, and messy references/ layout can slip through. This skill turns a ten-rule rubric into an executable review flow: it runs a deterministic lint first, then loads only the rubric file needed for the current rule, instead of stuffing all references into context.

How it works

Review mode accepts a skill directory or a SKILL.md path, runs a zero-dependency lint script, and gets P0–P3 counts, a verdict, and hard-fact findings. It then evaluates selected R1–R10 rubrics on demand and appends soft findings. Before editing, it asks whether to report only, fix P0/P1, or fix specific items. If approved, it uses targeted replacements and re-runs lint rather than rewriting freely.

Template mode confirms the skill name, purpose, trigger phrases, and whether it needs scripts or confirmation gates. It then scaffolds a new skill under .cursor/skills/<name> and applies constraints such as size limits, description quality, workflow checklist, argument parsing, and pre-delivery self-checks. The scaffold is verified with the linter, aiming for GOOD or OK_WITH_NOTES.

Boundaries

It reviews Skill structure, not arbitrary application code. The rule set is fixed: R1–R10 plus the lint script's hard rules. --quick runs lint only, --rules can narrow rubrics, and the target skill should not be modified before explicit confirmation.