/install oc-skill-audit
Skill Audit
Audit and score OpenClaw AgentSkills with a multi-dimensional scoring system.
Quick Start
"audit this skill" or "score path/to/SKILL.md"
Language Policy
- Audit reports: Always generate in two languages:
- User's language — for accessibility (detect from request language or target SKILL.md body language)
- English — for global shareability and comparability
- Both versions should be complete (scores, dimension analysis, improvement recommendations).
- Format: write the user-language version first, then the English version, separated by a clear divider.
- If the user already requested English, a single English version suffices (no duplication).
Scoring Overview
Total (0–100) = Weighted Average
Dimension Weight Score Range Description
─────────────────── ─────── ─────────── ───────────
A. Structure 20% 0–100 Spec compliance, frontmatter, file structure
B. Triggering 15% 0–100 description quality, keywords, triggering
C. Style Guide 20% 0–100 Judgment criteria, preservation/compression rules
D. Workflow 15% 0–100 Phase 0, sub-agent design
E. Sub-Agent Design 15% 0–100 Context injection, templates
F. Conciseness 15% 0–100 SKILL.md length, references separation
Grade: A(90+) | B(75+) | C(60+) | D(45+) | F(\x3C45)
How to Audit
- Read target SKILL.md and any references/
- Score each dimension using the rubric below
- Calculate weighted total and grade
- Report score card + improvement recommendations
Dimension A: Structure (20%)
Frontmatter, file structure, agentskills spec compliance.
| Check | Points | Criteria |
|---|---|---|
| Frontmatter exists | 15 | --- opening and closing present |
name valid |
10 | lowercase, hyphens only, 1-64 chars |
description present |
15 | Non-empty, 1-1024 chars |
description has keywords |
10 | Keywords in multiple languages |
| No auxiliary files | 10 | No README.md, CHANGELOG.md, etc. |
| references/ organized | 10 | If exists, 1-level deep, TOC for >100 lines |
| SKILL.md under 500 lines | 10 | Lean body, details in references/ |
| No duplication | 5 | Info in SKILL.md OR references, not both |
| Folder name matches name | 15 | skill-name/SKILL.md |
Max: 100
Dimension B: Triggering (15%)
Whether the description triggers the agent correctly.
| Check | Points | Criteria |
|---|---|---|
| "What" described | 20 | Clear description of what it does |
| "When" in description | 25 | Trigger conditions in description (not body) |
| Keywords (Korean) | 15 | Korean keywords included |
| Keywords (English) | 15 | English keywords included |
| Not too verbose | 10 | description ≤ 300 words (metadata cost) |
| Not too vague | 15 | Not overly broad (e.g., "helpful skill") |
Max: 100
Dimension C: Style Guide (20%)
Whether judgment criteria (preservation/compression) are specified.
Applicability:
- Applies: Text processing/transformation/summarization skills (tasks that modify original content)
- Partially applies: Code generation/modification skills (may need existing code preservation criteria)
- N/A: Structured data management (JSON/DB CRUD), simple tool wrappers, configuration management
For N/A skills, exclude this dimension and redistribute weights among remaining dimensions:
- Only C is N/A: A:24%, B:18%, D:18%, E:18%, F:22%
- Only E is N/A: A:24%, B:18%, C:24%, D:18%, F:16%
- Both C+E are N/A: A:30%, B:20%, D:20%, F:30%
| Check | Points | Criteria |
|---|---|---|
| Preservation criteria | 30 | "What to preserve" with specific examples |
| Compression criteria | 25 | "What to compress" with specific examples |
| Task essence defined | 20 | Core task definition (e.g., "summarization is restructuring, not compression") |
| Style rules explicit | 15 | Specific rules for tone, style, length, etc. |
| Anti-patterns | 10 | "What not to do" specified (bonus) |
Max: 100
Dimension D: Workflow (15%)
Whether Phase 0, execution order, and validation steps exist.
| Check | Points | Criteria |
|---|---|---|
| Phase 0 (analysis) first | 30 | Full understanding step before task execution |
| Clear phase numbering | 15 | Phase 0, 1, 2... structured |
| Context injection described | 25 | Explicit statement that Phase 0 results are injected into sub-agent prompts |
| Validation step | 20 | Result quality validation step (length/style/omissions) |
| Error handling | 10 | Failure response instructions |
Max: 100
Dimension E: Sub-Agent Design (15%)
Sub-agent prompt design quality.
Applicability:
- Applies: Skills that use sub-agents (parallel/sequential chunk processing, etc.)
- Partially applies: Skills that optionally mention sub-agent delegation (core is single-agent)
- N/A: Skills that do not use sub-agents at all
For N/A skills, exclude this dimension and redistribute weights among remaining dimensions:
- Only C is N/A: A:24%, B:18%, D:18%, E:18%, F:22%
- Only E is N/A: A:24%, B:18%, C:24%, D:18%, F:16%
- Both C+E are N/A: A:30%, B:20%, D:20%, F:30%
| Check | Points | Criteria |
|---|---|---|
| Prompt template exists | 20 | Template file in references/ |
[ ] placeholders |
20 | Blank fields filled by Phase 0 explicitly marked |
| Context fields defined | 20 | Full context, chunk position, preservation list |
| Style rules in template | 15 | 3-4 line style instructions included |
| Merge ≠ concatenate | 15 | Explicit prohibition of simple concatenation |
| Sequential pipeline | 10 | Sequential ordering for multi-stage output |
Max: 100
Dimension F: Conciseness (15%)
Context window efficiency.
| Check | Points | Criteria |
|---|---|---|
| SKILL.md body \x3C 500 lines | 40 | Length compliance |
| SKILL.md body \x3C 300 lines | 10 | Bonus (very concise) |
| No redundant instructions | 20 | No duplicate instructions |
| "Agent is smart" principle | 15 | Does not redundantly explain what the agent already knows |
| References linked | 15 | "When to read" description for each reference file |
Max: 100
Output Format
Delivery
Always do both:
- File: Save the full audit report (bilingual) to
~/.openclaw/workspace/skill-audit-reports/[skill-name]-audit-[YYYY-MM-DD-HHmm].md. If the same skill is audited again, a new timestamped file is created — never overwrite previous reports. Never save inside the target skill's directory — keep skill folders clean for distribution. - Response: Include a summary card in your reply — total score, grade, top 3 improvement recommendations, and file path. Do NOT paste the full report in the response.
Bilingual Output
Per the Language Policy above, generate two complete versions:
- User's language version (first)
- English version (second)
Separate them with a clear divider:
---
## English Version
If the user requested English, a single English version suffices.
Score Card
# Skill Audit: [skill-name]
**Version**: skill-audit v1.0
**Date**: YYYY-MM-DD
**Target**: path/to/SKILL.md
## Score: 78/100 (B)
| Dimension | Score | Weight | Weighted |
|-----------|-------|--------|----------|
| A. Structure | 85 | 20% | 17.0 |
| B. Triggering | 90 | 15% | 13.5 |
| C. Style Guide | 60 | 20% | 12.0 |
| D. Workflow | 70 | 15% | 10.5 |
| E. Sub-Agent Design | 80 | 15% | 12.0 |
| F. Conciseness | 85 | 15% | 12.8 |
| **Total** | | | **77.8** |
## Grade: B
---
## Dimension Details
### A. Structure (85/100)
✅ Frontmatter exists (+15)
✅ name valid: "summarize" (+10)
✅ description present (+15)
...
❌ SKILL.md over 500 lines (-10)
→ Tip: Split detailed specs into references/
### B. Triggering (90/100)
...
## Improvement Recommendations
1. **[High]** Add preservation/compression criteria to Style Guide → C score +20 expected
2. **[Medium]** Specify Phase 0 → D score +15 expected
3. **[Low]** Add references/ TOC → A score +5 expected
Improvement Priority Labels
- [Critical] — Skill may not function properly
- [High] — Major impact on quality
- [Medium] — Meaningful score improvement if addressed
- [Low] — Minor improvement
Batch Mode
Audit multiple skills at once:
"Audit all skills in the skills/ folder"
→ Generate individual score cards per skill, then produce a comparison table:
| Skill | Total | Grade | A | B | C | D | E | F |
|-------|-------|-------|---|---|---|---|---|---|
| summarize | 82 | B+ | 93 | 88 | 88 | 77 | 78 | 65 |
| changelog | 85 | B+ | 90 | 85 | 90 | 85 | N/A | 80 |
| scaffold | 81 | B | 85 | 80 | 75 | 80 | N/A | 85 |
References
references/scoring-rubric.md— Detailed scoring rubric with examples per dimensionexamples/— Completed audit reports and batch comparisons for validated skills
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install oc-skill-audit - After installation, invoke the skill by name or use
/oc-skill-audit - Provide required inputs per the skill's parameter spec and get structured output
What is Skill Audit?
Audit and score OpenClaw AgentSkills against structural compliance, quality standards, and OpenClaw-specific architecture patterns. Produces a 0-100 score wi... It is an AI Agent Skill for Claude Code / OpenClaw, with 59 downloads so far.
How do I install Skill Audit?
Run "/install oc-skill-audit" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is Skill Audit free?
Yes, Skill Audit is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does Skill Audit support?
Skill Audit is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created Skill Audit?
It is built and maintained by haseo-ai (@haseo-ai); the current version is v1.0.0.