Review the specified Skill's specification compliance and output a structured report. Do not auto-fix; only report problems.
Step 1 — Load context
- Parse
$ARGUMENTS:- If it is a directory path → read the
SKILL.md - If it is a file path → read that file directly
- If the path does not exist or there is no SKILL.md → output an error message and terminate
- If it is a directory path → read the
- Use Glob to list all files under the Skill directory (including subdirectories)
- Read the official specification:
- Under this Skill directory,
references/The-Agent-Skills-Specification.md(Agent Skills official specification) - Under this Skill directory,
references/Extend-Claude-with-skills.md(Claude Code official documentation) - Under this Skill directory,
references/The-Complete-Guide-to-Building-Skill-for-Claude.md(Complete Claude Skill build guide) - If neither exists → complete the review based on built-in knowledge and begin the report output with the following warning block:
⚠️ Confidence: Medium — Official specification documents not loaded (path does not exist); results are based on
references/checklist.mdand built-in knowledge. It is recommended to update the document path and review again.
- Under this Skill directory,
- Read the complete checklist: under this Skill directory,
references/checklist.md
Step 2 — Check item by item
Judge item by item against the checklist in references/checklist.md . Mark each item ✅ Pass, ❌ Fail, ⚠️ Suggested optimization.
Use Grep to assist in judging the following items (to avoid omissions from pure visual inspection):
- A3/C1 Tool declarations:Grep
\bBash\b、\bRead\b、\bWrite\b、\bEdit\b、\bGlob\b、\bGrep\b、\bWebFetch\b、\bSkill\b、\bTask\bCheck where tools actually appear and cross-compare withallowed-toolsCross-comparison - A11 Unknown fields: Grep the frontmatter block (
^---extract all keys between them and compare with the list of legal fields - D1/D2 File references:Grep
\.md、.py、.shCheck paths with extensions such as these, and cross-check with the Glob file list for orphaned/broken-link files - E1 Non-standard placeholders:Grep
\{\{and<[A-Z_]+>detect non-standard syntax such as{{...}}and<PLACEHOLDER>and other non-standard syntax - F1 Hardcoded paths:Grep
/Users/、/home/、/root/absolute path patterns such as these
Checks cover 6 categories and 33 items in total:
| Category | Number range | Check points |
|---|---|---|
| A. Frontmatter compliance | A1-A13 | name/description format and content, field validity, security restrictions, directory hygiene |
| B. Instruction content quality | B1-B6 | Step completeness, branch coverage, exit conditions, output format, line count control |
| C. Tool usage compliance | C1-C3 | Declaration consistency, parameter constraints, Task subagents |
| D. File reference validity | D1-D3 | Forward/reverse reference checks, path format |
| E. Variables and dynamic content | E1-E5 | Standard syntax, index reasonableness, dynamic injection security, argument-hint consistency |
| F. Portability | F1-F3 | Hardcoded paths, language/culture binding, toolchain hard binding |
Judgment principles:
- Items with explicit rules in the specification document → strict determination (❌ or ✅)
- Items that are best practice recommendations → mark as ⚠️ rather than ❌
- When uncertain whether it is a violation → mark as ⚠️ and explain the doubt
Step 3 — Output report
When problems are found
Sort in descending order of severity:
| # | Check item | Severity | Location | Problem | Suggested fix |
|---|
Severity definitions:
- High: Skill cannot be triggered, execution errors occur, or incorrect results are produced
- Medium: Does not block execution but reduces reliability or maintainability
- Low: Best practice recommendation
Suggested fix requirements:
- High/medium severity items: must provide directly usable modification content (such as corrected YAML fragments, revised text)
- Low severity items: explaining the improvement direction is sufficient
When there are no problems
Output:✅ 未发现问题。已检查 A1-A13、B1-B6、C1-C3、D1-D3、E1-E5、F1-F3 共 33 项。
Example
Example 1: Typical problem report
Input:/review-skill .claude/skills/my-deploy/
Output fragment:
| # | Check item | Severity | Location | Problem | Suggested fix |
|---|---|---|---|---|---|
| A2 | description |
Medium | frontmatter | Only 15 characters "deploy application", lacking trigger words and usage timing | Expand to:description: 部署应用到生产环境。Use when asked to deploy, release, or push to production. |
| A4 | disable-model-invocation |
Medium | frontmatter | Contains deployment operations but is not set to true | Add disable-model-invocation: true |
| C1 | Declaration consistency | High | Step 3 | Bash is used but not declared in allowed-tools | Add to allowed-tools Bash(deploy *) |
Example 2: All passed
Input:/review-skill .claude/skills/explain-code/
Output:✅ 未发现问题。已检查 A1-A13、B1-B6、C1-C3、D1-D3、E1-E5、F1-F3 共 33 项。
Handling when the determination is ambiguous
Some check items (such as B1 step chain, B5 behavior impact) require subjective judgment. In this case, mark them as ⚠️ and explain the doubt in detail in the "Problem" column for the user to judge on their own.