Skill Self Check

Skill Self-Check

Run a fast static review with deterministic gate status and ranked fixes. Full checks: CHECKLIST.md.

Workflow prompt audit: N/A — one agent instruction context; no separate model calls.

• scripts/hard_gates.py exclusively owns package_health, gate_verdict, script findings, severity, and exit code.
• scripts/verify_fix.py exclusively owns fix verification against a saved baseline.
• Script scores are informational only.
• Qualitative review is optional and model-owned; script findings, gate status, counts, severity, and exit code remain unchanged.
• Edit the target only after explicit user authorization.

When to use

• Review an existing Skill before controlled use, installation, or sharing.
• Diagnose deterministic blockers and provide paste-ready fixes.
• Recheck user-authorized repairs against a saved baseline.

When NOT to use

• Create a Skill from scratch.
• Run behavioral evaluations or multi-agent smoke tests.
• Generate profiles, presentations, or visual scorecards.

Check axes

This audit always reports on:

• Package health preflight — one installable root, name/root alignment, standard top-level directories, portable paths, valid resource references, filename/residue hygiene, duplicate resources, and static installability (script; blocks maturity assessment)
• Hard structure — frontmatter, name, description shape (script)
• Explicit gate — valid package, named required checks, and no script Criticals; independent of numeric scores (script; blocking)
• Basic usable score — 0–5 informational diagnostic (script)
• Contract clarity score — 0–5 informational diagnostic (script)
• Support kit score — references / examples / memory / scripts; N/A allowed; informational only (script)
• Predictability — completion criteria, no-op, negation, sprawl (model + script hints)
• Anatomy — workflow quality, rationalizations (model + script hints)
• PDCA loop — Plan / Do / Check / Act all explicit (model; see references)
• SMART outcomes — Specific, Measurable, Achievable, Relevant, run-bound exit (model)
• Token consumption — static SKILL.md input estimate with a recommended budget ceiling, replaced by trusted input/output/total usage when behavior evidence supplies it (script)
• Runtime duration — target execution time only from trusted behavior evidence; otherwise not_measured (script)
• Loop guard — every loop/retry instruction carries a stop condition (max attempts / timeout / escalate); open-ended refinement requires a guard before the Skill ships (script EFF.*)

Inputs

1. Locate the target skill directory (must contain SKILL.md).
2. Prefer an explicit path from the user.
3. If missing: ask once, or use the skill they just created/opened in this conversation.

Completion criterion: You know the absolute path to the target skill directory.

Process

Default — fast hard-gate audit

Run the bundled checker directly. This Skill operates independently of the other shipped Skills:

python scripts/hard_gates.py /absolute/path/to/target-skill --pretty

Windows may use py -3.

For multi-Skill packs, pass --repo-root /absolute/repository; otherwise links stay target-local. It allows only in-root relative links; reports label each resolution_scope as target or repo. Absolute/escaping paths stay Critical.

If fixes may follow, save the same run as a baseline outside the target and source repository:

python scripts/hard_gates.py /absolute/path/to/target-skill \
  --out-json /private/path/baseline.json --pretty

• Parse stdout JSON. Treat stderr as a human summary only.
• Read package_health, gate_verdict, gate_reasons, and findings first.
• Exit code 1 means the gate did not pass; report the fixes normally.
• Treat scores and operational metrics as informational. Use gate_verdict as the exclusive gate authority instead of scores.ship_floor_met.
• Present maturity scores as a package assessment only when package health is valid and assessable.

Completion criterion: JSON parsed; package_health, gate_verdict, gate_reasons, and findings are available.

Rank and explain deterministic findings

Map script findings with severity: critical|should_fix|nice into the report.
Explain and suggest rewrites; keep every script Critical in failed status. Use references/plain-language-response.md to translate the source result without creating a second audit.

PKG.* and EFF.* are mechanical: the fix does not depend on the user's business, so write the rewrite yourself using references/fix-templates.md instead of asking.

1. Gate verdict and plain-language reasons.
2. Every script Critical, each with 问题 → 为什么 → 可直接采用的建议改法.
3. At most three highest-priority Should fix findings.
4. One next action: say 「按意见改」 to authorize edits, or explicitly ask for deep audit / full static audit.

Fast-mode completion criterion: Every script Critical is covered, no more than three Should fix items are shown, and each displayed finding has an actionable fix. Stop here unless the user explicitly requests another route.

Optional routes

Load only the reference selected by the user's explicit request:

For an explicitly requested deep/full technical report, use REPORT-TEMPLATE.md only for that explicit route.

Verification

• hard_gates.py was executed on the target directory
• gate_verdict and gate_reasons were read before deprecated score fields
• Every script Critical has a paste-ready fix
• No more than three Should fix items appear in the default response
• Token consumption states estimated, observed, or not_assessed with scope
• Runtime duration is observed only with trusted behavior evidence; otherwise not_measured
• No script Critical was overridden
• User was advised whether the deterministic gate passed

Only for explicitly requested routes:

• Deep audit: model findings are advisory and labeled source: model_review
• Deep audit: PDCA×SMART matrix gaps map to advisory priorities
• Prompt optimization: static reduction, quality, and behavior claims remain separate
• Prompt optimization: comparative claims use a saved pre-edit baseline
• Workflow prompt audit: node findings remain separate from gate_verdict
• Full static audit: source JSON is outside the target and source repository
• Full static audit: target fingerprint is unchanged
• Applied fixes: verify_fix.py ran against the pre-fix baseline
• Applied fixes: every introduced finding is fixed or explicitly reported

Common Rationalizations

Red Flags

• Writing a gate verdict without running the script
• Running PDCA×SMART or the full static audit without an explicit request
• Letting model review change gate_verdict, Critical counts, or exit status
• Inventing exclusions, triggers, or acceptance evidence
• Claiming findings are fixed without a verify_fix.py before/after table
• Editing the target before the user explicitly authorizes it

Out of scope

• Creating a skill from scratch
• Automated multi-case behavioral evals (v2)
• Editing the target unless the user explicitly asks
• Inventing quarterly OKRs for a skill that only needs a session exit criterion