Audit and validate existing Claude Code skills for quality, triggering accuracy, structure compliance, and best practices. Scores skills on a 0-100 scale and provides prioritized improvement recommendations. Use when user says "review skill", "audit skill", "check skill", "validate skill", or "skill quality".
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Audit and validate existing Claude Code skills for quality, triggering accuracy, structure compliance, and best practices. Scores skills on a 0-100 scale and provides prioritized improvement recommendations. Use when user says "review skill", "audit skill", "check skill", "validate skill", or "skill quality".
Skill Review & Validation
Process
Step 1: Locate Skill Files
Accept input as:
Path to a skill directory
Skill name (search in ~/.claude/skills/)
URL to a GitHub repository
Read all .md files, scripts, and asset files.
Step 2: Structure Validation
Run python scripts/validate_skill.py <path> for programmatic checks.
Manual verification:
SKILL.md exists (exact case)
No README.md inside skill folder
Folder name matches name field
Valid kebab-case naming (1-64 chars)
No "claude" or "anthropic" in name
Step 3: Frontmatter Audit
Check
Pass Criteria
Name format
kebab-case, 1-64 chars, no leading/trailing hyphens
Description present
Non-empty, 1-1024 characters
Description has WHAT
Explains capabilities
Description has WHEN
Includes trigger phrases
Description has keywords
Domain-specific terms included
No XML tags
No < or > characters
Optional fields valid
license, compatibility (<500 chars), metadata
Step 4: Triggering Analysis
Assess the description for activation quality:
Under-triggering risks:
Too generic ("Helps with projects")
Missing common paraphrases
No domain keywords
Missing file type mentions (if relevant)
Over-triggering risks:
Too broad ("Processes documents")
Overlaps with built-in Claude capabilities
Missing negative triggers for disambiguation
Generate test queries:
5 queries that SHOULD trigger the skill
5 queries that SHOULD NOT trigger
3 edge cases (ambiguous queries)
Step 5: Instruction Quality
Criterion
Score (0-10)
Specificity
Are instructions actionable? (not "validate properly")
Completeness
All workflows covered?
Error handling
Common failures addressed?
Examples
Concrete examples provided?
Progressive disclosure
Detailed docs in references/ not SKILL.md?
Length
Under 500 lines / 5000 tokens?
Cross-references
Clear links to references/scripts?
Step 6: Architecture Review (Multi-skill)
For skills with sub-skills:
Main skill has clear routing table
Sub-skills have focused responsibilities
Cross-references are valid (files exist)
Naming follows parent-child convention
Shared references in parent, not duplicated
Agents have clear roles (if Tier 4)
Step 7: Script Quality (if present)
Docstrings with purpose, input, output
CLI interface (argparse or similar)
Structured output (JSON)
Error handling (try/except with clear messages)
No hardcoded paths or secrets
Minimal dependencies
Step 8: Generate Skill Health Score
Scoring methodology (0-100):
Category
Weight
Checks
Frontmatter Quality
25%
Name, description, format
Trigger Accuracy
20%
WHAT + WHEN + keywords
Instruction Quality
25%
Specificity, completeness, examples
Structure Compliance
15%
File naming, organization, references
Script Quality
10%
If applicable (full marks if no scripts needed)
Progressive Disclosure
5%
Proper use of 3-level system
Step 9: Generate Trigger Eval Set
After reviewing, generate a structured trigger eval set for ongoing testing:
Run python scripts/generate_eval_set.py <path> to auto-generate a starter set
Review and refine the generated queries:
Ensure 8-10 should-trigger queries cover different phrasings and edge cases
Ensure 8-10 should-not-trigger queries are near-misses (not obviously irrelevant)
Include casual speech, typos, and uncommon domain uses in should-trigger set
Save the eval set to evals/evals.json in the skill directory
Good queries are realistic and specific (include file paths, context, domain details).
Bad queries are overly generic ("format this data") or obviously irrelevant.
Run python scripts/optimize_description.py <path> --eval-set evals/evals.json
to score the current description and get improvement suggestions
Recommend running /skill-forge eval <path> for full functional evaluation