← ClaudeAtlas

review-skilllisted

Reviews and audits an existing skill, then tightens it — a fast, static, read-only audit of a `SKILL.md`'s triggering, scope, structure, prose, and domain accuracy, with severity-ranked findings applied on approval. Use whenever the user wants to review, critique, audit, lint, tighten, or improve a skill they already have, fine-tune a skill drafted with skill-creator, or asks why a skill is too verbose, won't trigger, or feels off — even when they just paste a SKILL.md and ask for feedback. Defers the empirical loop (running evals, benchmarking, the automated description optimizer, packaging) to skill-creator; this is the static counterpart that reads rather than measures.
azborgonovo/ai-skills · ★ 0 · Code & Development · score 75
Install: claude install-skill azborgonovo/ai-skills
# Review Skill Audit an existing skill against how good skills actually behave, then tighten it — a close reading that surfaces what undermines the skill and fixes it on the user's say-so. When the user wants measured triggering accuracy or eval-graded iteration, hand off to skill-creator, which owns the empirical description-optimizing loop; don't rebuild that loop here. Work in two phases: **critique first, edit second.** Never rewrite the skill before the user has seen the findings and chosen what to apply. ## Resolve the target Take the skill from `$ARGUMENTS` — a skill name (resolve to `skills/<name>/SKILL.md` or the installed `~/.claude/skills/<name>/SKILL.md`) or a direct path. With no argument, review the skill in progress: the `SKILL.md` most recently edited this session or showing in `git status`; if that's ambiguous, confirm which one before reviewing. Read the whole bundle, not just `SKILL.md` — note any `scripts/`, `references/`, `assets/`, and `evals/`, since their presence (or absence) is itself a finding. ## Load the lenses Two lenses sharpen the review when available. Use each only if present, and do not reconstruct it inline when it's missing — a partial review beats a duplicated rubric that drifts out of sync. - **Project conventions** — if the repo has a `CLAUDE.md` with skill-authoring guidance, read it and treat its rules as binding for this review (this repo's, for instance, bans `---` section separators, reserves bold for lead-ins and directives