Menü
Use when auditing or improving an agent skill package before it ships or is handed off.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Basierend auf der SOC-Berufsklassifikation
Use when the user explicitly runs /setup-precommit to install or refresh a quiet, staged-only git pre-commit hook (secretlint → oxlint → oxfmt) wired through package.json "prepare". User-invoked command only, never auto-triggered.
Use when the user asks to create, make, or start a new branch (e.g. `/branch`), or wants a conventional-commit-style branch name generated from the current work.
Use when the user is working on their own PR — drafting/creating it, handling reviewer comments, verifying fixes, or requesting re-review.
Use when reviewing someone else's PR at any stage of the review lifecycle — digest a PR, write a review, or verify follow-up changes.
Use when fixing a bug, refactoring existing code without changing its behavior, or adding a feature that extends existing behavior — any code change that needs a verifiable foundation before the change is made.
Use when the user wants to A/B test prompts or workflows, validate behavioral equivalence, or stress-test non-deterministic outputs at scale via repeated multi-agent runs.
| name | evaluate-skills |
| description | Use when auditing or improving an agent skill package before it ships or is handed off. |
| argument-hint | [skill dir or SKILL.md path; one or more] |
| disable-model-invocation | true |
| allowed-tools | Read Grep Glob Bash(node:*) |
Audit a TARGET agent skill package against the skill-evaluation rubric and report the findings with proposed fixes. This is an evaluation tool — it inspects and reports on another skill; it does not apply changes or run that skill's workflow.
Resolve every bundled path below against this skill's own directory.
Work the steps in order. Each step's prompt is its own file under steps/ — read it when
you reach that step, do what it says, and honor any halt condition it states before moving
on. Do not skip steps or read ahead.
steps/1-target-resolution.mdsteps/2-preflight.md · classify deterministically; carry findings into the report. Halt only for a target that cannot be read or parsed (review is impossible)steps/3-read-target.mdsteps/4-judgment-review.mdsteps/5-output.md