Skill Review

Language

Match user's language: Respond in the same language the user uses.

Overview

Review an agent skill through three layers: automated linting (hard rules), contextual finding evaluation (agent judges with context), and structured semantic review (deep analysis against best practices). Optionally fix issues and verify fixes through independent subagent validation.

Workflow

Progress:

Step 1: Identify the skill
Step 2: Review (linting + profile + findings + semantic)
Step 3: Present findings
Step 4: Fix gate
Step 5: Verify & iterate

Step 1: Identify the Skill

Accept the skill location as a directory path containing SKILL.md. Auto-detect if the current working directory contains one.

Step 2: Review

Run both tools, then evaluate.

2a. Automated linting:

python3 {SKILL_DIR}/scripts/validate.py run --path <skill-path>

Output: checks (hard-rule verdicts → linter grade) + findings (soft detections needing judgment).

2b. Profile extraction:

bash {SKILL_DIR}/scripts/analyze.sh analyze <skill-path>

Note skill level (l0/l0plus/l1) and feature flags.

2c. Contextual findings review:

Judge each finding using its context_hint. See references/semantic_dimensions.md § Finding Judgment Table for decision criteria. Promote real issues to warnings; dismiss the rest with a brief note.

2d. Semantic review:

Read the skill's SKILL.md fully, then score each dimension (0-3). Read references/semantic_dimensions.md for the full checklist per dimension:

Dimension	What it evaluates
5.1 Description Quality	Trigger phrases, length, voice, specificity
5.2 Workflow Design	Steps, decision points, SKILL.md length
5.3 Runtime Robustness	Preflight, degradation, troubleshooting (L0+/L1 only)
5.4 Script Quality	JSON output, error handling, exit codes (scripts only)
5.5 UX Practices	Language, checklist, completion report (applicability matrix)
5.6 Setup Flow Integrity	Bootstrap safety, live validation, credential security

Step 3: Present Findings

Format the report:

[Skill Review] <skill-name>

═══ Linter ═══
Grade: <letter> (<pass>/<total> passed, <warn> warnings, <fail> failures)

Failures:
  ✗ <check_id>: <message> → Fix: <fix>

Warnings:
  ⚠ <check_id>: <message>

Findings (agent-reviewed):
  ✓ <finding_id>: dismissed — <reason>
  ⚠ <finding_id>: promoted to warning — <reason>

═══ Semantic Review ═══
5.1 Description Quality:   <score>/3  <one-line assessment>
5.2 Workflow Design:        <score>/3  <one-line assessment>
5.3 Runtime Robustness:     <score>/3  <one-line assessment>
5.4 Script Quality:         <score>/3  <one-line assessment>
5.5 UX Practices:           <score>/3  <one-line assessment>
5.6 Setup Flow Integrity:   <score>/3  <one-line assessment>
                            ─────────
Semantic Score:             <total>/18

═══ Improvement Suggestions ═══
For each dimension scoring < 3, provide:
  1. What to change and why
  2. Which file to edit
  3. A concrete before/after example or specific instruction
  4. Priority: High (functionality/UX) / Medium (convention) / Low (polish)

If linter grade is A and semantic score ≥ 15: congratulate and suggest publishing with better-skill-publish.

Step 4: Fix Gate

Ask the user:

Fix all — Apply all suggested changes, then auto-verify (Step 5)
Pick and choose — Select specific items, then auto-verify (Step 5)
None — End here, use the report as reference

After fixing, do NOT self-evaluate. Proceed directly to Step 5.

Step 5: Verify & Iterate

CRITICAL: Verification must be independent. The agent that fixed cannot judge its own work.

5a. Dispatch verification subagent:

Agent(subagent_type: "general-purpose", description: "Verify skill review", prompt: <template>)

Use the prompt template from references/semantic_dimensions.md § Verification Subagent Prompt Template. The verification subagent must:

Run validate.py and analyze.sh fresh (never trust cached results)
Read references/semantic_dimensions.md for scoring criteria
Score all 6 dimensions independently, citing file:line evidence
Return linter grade + per-dimension scores + specific FAIL items

5b. Process verification result:

All pass → Report success, done
Has failures → Fix ONLY the specific FAIL items from the verification report, then dispatch a NEW subagent to verify again
Max 3 rounds — if issues persist after 3 fix-verify cycles, report remaining issues and let the user decide

Anti-patterns (never do these):

Anti-pattern	Why it's wrong	Correct approach
Self-evaluate after fixing	Blind to own mistakes	Always dispatch subagent
Reuse same subagent for re-verify	Already biased by prior assessment	New subagent each round
Fix without running tools first	Subjective judgment misses issues	Tools produce objective data
Ignore verification report	Discards independent evidence	Fix only reported FAIL items

Linter Reference

Grading (hard-rule checks only):

A — All pass, zero warnings
B — All pass, some warnings
C — 1–2 failures
D — 3+ failures
F — SKILL.md missing or no valid frontmatter

Check categories:

Category	Checks
structure	SKILL.md exists, frontmatter, required fields, directory layout
naming	Kebab-case, length, no consecutive hyphens, matches directory
content	Description length, body length, heading structure
paths	Referenced files exist, scripts executable
security	No secrets, no template placeholders

References

references/semantic_dimensions.md — Full checklist for each review dimension + finding judgment table + verification prompt template
references/validation_rules.md — Rationale for each linter check
references/improvement_patterns.md — Knowledge base of improvement patterns with examples
references/best_practices.md — Skill design conventions and quick reference

better-skill-review