openclaw-skill-eeta-audit（OpenClaw 技能 eeta 审核）

EEAT OpenClaw Skill Audit

AI Agent Skill Quality Assurance — This skill adapts the CORE-EEAT framework to evaluate OpenClaw Skills, ensuring they deliver meaningful utility while maintaining security and reliability.

Skill Overview

OpenClaw Skills are modular capability extensions for AI agents, defined by SKILL.md files with YAML frontmatter and prompt instructions. This skill evaluates skill quality through 80 standardized criteria across 8 core dimensions, generating comprehensive audit reports including utility scores, security assessments, and actionable improvement recommendations.

Core Transformation:

• From: Install Skills blindly → Hope they work
• To: Systematic vetting → Data-driven skill selection

OpenClaw Skill Structure

Every OpenClaw Skill consists of:

my-skill/
├── SKILL.md          # Core definition (YAML + Markdown instructions)
├── scripts/          # Optional executable scripts
│   └── main.py
└── references/       # Optional configuration and resources
    └── config.json

Key Components:

• YAML Frontmatter: Skill metadata (name, description, version, dependencies, gates)
• Prompt Instructions: How the AI should use this skill
• Scripts: Optional executable code for complex operations
• Gating Mechanism: Conditional activation (bins, env, os checks)

Applicable Scenarios

Use this skill when users request:

Skill Selection

• "Evaluate this skill before installing"
• "Compare two skills for the same task"
• "Which skill should I use for X?"

Security Vetting

• "Is this skill safe to run?"
• "Scan for security vulnerabilities"
• "Check permission boundaries"

Skill Development

• "Audit my skill for quality issues"
• "How to improve my skill's documentation?"
• "What security best practices am I missing?"

Skill Maintenance

• "Review installed skills for quality"
• "Identify deprecated or risky skills"
• "Prioritize skill updates"

Core Capabilities

This skill can:

1. Complete 80-Item Audit: Score each CORE-EEAT item adapted for OpenClaw Skills

1. Utility Scoring: Evaluate task completion efficiency and comparative value

1. Security Assessment: Three-level security evaluation (Pass/Caution/Risk)

1. Gating Validation: Check conditional activation requirements (bins, env, os)

1. Veto Item Detection: Flag critical security violations (command injection, data leakage)

1. Priority Ranking: Identify top 5 improvements by impact

1. Comparative Analysis: Compare skills for same use case

Skill Categories

This skill supports 6 OpenClaw Skill types, each with different evaluation priorities:

Productivity Skills

• Definition: Gmail, Calendar, Google Drive, Microsoft Office integration
• Focus: Task completion accuracy, API reliability, error handling
• Weights: C: 30% | R: 25% | Exp: 20% | Ept: 15% | O: 5% | E: 0% | A: 5% | T: 0%

Development Skills

• Definition: Code generation, debugging, GitHub automation, CI/CD
• Focus: Code quality, correctness, testing, security best practices
• Weights: C: 25% | O: 20% | R: 20% | Ept: 20% | E: 5% | Exp: 5% | A: 5% | T: 0%

Research Skills

• Definition: Web search, web fetch, document summarization, data analysis
• Focus: Information accuracy, source credibility, citation quality
• Weights: C: 25% | R: 25% | A: 20% | E: 15% | O: 10% | Exp: 0% | Ept: 0% | T: 5%

Automation Skills

• Definition: Browser automation, file operations, shell commands, task scheduling
• Focus: Security, error handling, robustness, permissions
• Weights: T: 30% | C: 25% | R: 20% | O: 15% | Exp: 5% | Ept: 5% | E: 0% | A: 0%

Content Skills

• Definition: Text generation, translation, image generation, audio processing
• Focus: Output quality, style consistency, creative value
• Weights: C: 30% | E: 25% | Exp: 20% | O: 15% | Ept: 5% | R: 5% | A: 0% | T: 0%

System Skills

• Definition: System monitoring, resource management, network tools, debugging
• Focus: Performance, reliability, security, compatibility
• Weights: T: 25% | C: 20% | R: 20% | E: 15% | O: 10% | Ept: 5% | Exp: 5% | A: 0%

8 Progressive Quality Gates

Gate 1: Metadata Validation (Pre-Installation)

When: Before installing any skill

Duration: 2-5 minutes

Items:

• C01: YAML frontmatter present and valid
• C02: Skill name and description clear
• O01: Skill structure follows OpenClaw convention
• T04: No suspicious dependencies or permissions

Deliverable: Metadata Validation Report

Failure: Do not install. Contact skill author or fix manually.

Gate 2: Gating Mechanism Check

When: After metadata validation, before activation

Duration: 1-2 minutes

Items:

• O02: Required tools exist (bins check)
• O03: Required environment variables set (env check)
• O04: OS compatibility verified (os check)
• T07: No conflicting permissions

Deliverable: Gating Compatibility Report

Failure: Skill will not activate. Fix environment or choose alternative.

Gate 3: Security Pre-Check

When: Before first execution

Duration: 3-5 minutes

Items:

• T01: No command injection vulnerabilities
• T02: No data leakage risks
• T03: Input validation present
• T04: Permissions are minimal principle

Deliverable: Security Pre-Check Report

Failure: Do not execute. Review code or choose alternative.

Gate 4: Prompt Quality Review

When: During skill development or installation

Duration: 5-10 minutes

Items:

• C03: Instructions are clear and actionable
• C04: Tool usage patterns are explicit
• Ept03: Professional terminology used
• Ept05: Follows prompt engineering best practices

Deliverable: Prompt Quality Report

Failure: Skill may misbehave. Refine SKILL.md instructions.

Gate 5: Script Security Audit

When: For skills with executable scripts

Duration: 10-20 minutes

Items:

• T05: No hardcoded secrets or API keys
• T06: No eval/exec of user input
• T07: Proper error handling and logging
• T08: Resource cleanup implemented

Deliverable: Script Security Audit

Failure: Security risk. Audit scripts or avoid skill.

Gate 6: Utility Testing

When: During skill evaluation

Duration: 15-30 minutes

Items:

• C05: Completes intended tasks successfully
• C06: Output quality meets expectations
• Exp01: Usage examples provided and working
• Exp02: Performance characteristics documented

Deliverable: Utility Test Report

Failure: Skill doesn't deliver value. Consider alternatives.

Gate 7: Comparative Analysis

When: Selecting between multiple skills for same task

Duration: 10-15 minutes

Items:

• E01: Offers unique value vs alternatives
• E02: Better performance or efficiency
• A01: Active maintenance and updates
• A02: Community adoption and reviews

Deliverable: Comparative Analysis Report

Failure: Alternative skill may be better. Choose based on scores.

Gate 8: Full 80-Item Audit

When: Critical skills, major updates, security reviews

Duration: 30-60 minutes

Items: All 80 items across all 8 dimensions

Deliverable: Comprehensive EEAT Audit Report

Failure: Address critical issues before production use.

Usage Examples

Example 1: Pre-Installation Security Check

User Request:

Check if this skill is safe to install:
clawhub install agent-browser

Response:

## EEAT Skill Security Audit: agent-browser

### Gate 1: Metadata Validation ✅
- [x] YAML frontmatter valid
- [x] Description clear: "Browser automation for web scraping"
- [x] Version: 1.2.3
- [x] Dependencies: playwright (verified safe)

### Gate 2: Gating Check ⚠️
- [x] Bins: playwright installed
- [x] Env: PLAYWRIGHT_BROWSERS_PATH set
- [ ] OS: macOS/Windows only (Linux limited support)

### Gate 3: Security Pre-Check ✅
- [x] T01: No command injection detected
- [x] T02: No data leakage risks
- [x] T03: Input validation on URLs
- [x] T04: Minimal permissions (browser control only)

### Security Score: 92/100
**Security Status**: CAUTION ⚠️
- Linux support limited
- Requires browser installation (system-level change)

### Recommendation
✅ **Install with Caution**
- Safe for macOS/Windows users
- Linux users should check compatibility
- Review permissions before first use

Example 2: Compare Two Skills

User Request:

Compare tavily-search vs web-search for web search

Response:

## Skill Comparative Analysis: Web Search

### Skill A: tavily-search

**Utility Score**: 85/100
- C05: High accuracy search results ✅
- C06: Clean, structured output ✅
- Exp01: Excellent usage examples ✅
- E01: AI-optimized for agents 🎯

**Security Score**: 95/100
- T01-T04: All checks pass ✅
- T07: No external dependencies ✅

**Overall**: 88/100 (Very Good)

### Skill B: web-search (bundled)

**Utility Score**: 70/100
- C05: Basic search functionality ✅
- C06: Output quality varies ⚠️
- Exp01: Limited examples ⚠️
- E01: Generic implementation ⚠️

**Security Score**: 90/100
- T01-T04: Safe but basic ✅
- T07: Uses external APIs ⚠️

**Overall**: 75/100 (Good)

### Recommendation
✅ **Choose tavily-search**
- +18 points higher utility
- AI-optimized for agent workflows
- Better structured output for downstream processing
- Higher security score

**Trade-off**: Requires API key setup vs. no setup for web-search

Example 3: Full Skill Audit

User Request:

Audit my custom skill: email-automation
[Paste SKILL.md content]

Response:

## EEAT Skill Audit: email-automation

### Overview
- **Skill Type**: Productivity
- **Total Score**: 62/100 (Fair)
- **Utility Score**: 75/100 (Good)
- **Security Score**: 45/100 (CAUTION ⚠️)
- **Veto Item**: ⚠️ T02 Triggered (Data Leakage Risk)

### Dimension Scores

| Dimension | Score | Rating | Weight | Weighted Score |
|-----------|-------|--------|--------|----------------|
| C — Correctness | 70/100 | Good | 30% | 21 |
| O — Organization | 65/100 | Fair | 5% | 3.25 |
| R — Referenceability | 50/100 | Fair | 25% | 12.5 |
| Exp — Experience | 80/100 | Good | 20% | 16 |
| Ept — Expertise | 55/100 | Fair | 15% | 8.25 |
| E — Exclusivity | 40/100 | Poor | 0% | 0 |
| A — Authority | 60/100 | Fair | 5% | 3 |
| T — Trust | 45/100 | Poor | 0% | 0 |
| **Weighted Total** | | | | **64** |

### Critical Issues (Veto Items)

⚠️ **T02: Data Leakage Risk**
**Issue**: Skill stores API credentials in plain text in SKILL.md

SKILL.md

credentials: smtp_password: "mypassword123" # ⚠️ SECURITY RISK

**Action**: Move credentials to environment variables

credentials: smtp_password: "${SMTP_PASSWORD}" # ✅ SECURE

### Top 5 Priority Improvements

1. **T02 Data Leakage** — Remove hardcoded credentials
   - Current: Fail | Potential Gain: 8 weighted points
   - Action: Use environment variables for all secrets

2. **R02 Coverage** — Add error handling examples
   - Current: Fail | Potential Gain: 6.25 weighted points
   - Action: Document error scenarios and recovery

3. **Ept01 Documentation** — Improve prompt instructions
   - Current: Partial | Potential Gain: 4.5 weighted points
   - Action: Add step-by-step usage examples

4. **R03 Source Authority** — Verify email library security
   - Current: Partial | Potential Gain: 3.75 weighted points
   - Action: Audit nodemailer dependency for vulnerabilities

5. **O01 Structure** — Add scripts/ directory for complex logic
   - Current: Partial | Potential Gain: 2.5 weighted points
   - Action: Move complex operations to Python scripts

### Action Plan

#### Quick Wins (Fix immediately)
- [ ] Move all credentials to environment variables
- [ ] Add error handling documentation

#### Medium Investment (This week)
- [ ] Add comprehensive usage examples
- [ ] Implement proper logging in scripts

#### Strategic (Next sprint)
- [ ] Add test suite with edge cases
- [ ] Implement retry logic for failed sends
- [ ] Add HTML email support

### Recommendation
⚠️ **Do Not Install Until Fixed**
- Security risk (T02 veto) must be addressed
- After fixes, expected score: 78/100 (Good)

Reference Documents

• references/openclaw-skill-benchmark.md — Complete 80-item benchmark adapted for OpenClaw Skills
• references/skill-security-checklist.md — Security-specific evaluation criteria
• references/utility-testing-guide.md — How to test skill utility and comparative value
• workflow-optimization-analysis.md — Adaptation strategy from code to skills

Key Differences: Code vs. Skill Audit

Success Points

1. Security-First Approach — OpenClaw Skills have system-level access; security is non-negotiable

1. Comparative Utility — Evaluate skills relative to baseline, not in isolation

1. Gating Validation — Ensure skills only activate when dependencies are met

1. Prompt Quality — SKILL.md instructions determine skill behavior; quality matters

1. Minimal Permissions — Skills should only request necessary access

1. Active Maintenance — Prioritize skills with recent updates and community support

1. Real-World Testing — Test with actual use cases, not synthetic scenarios

Optimization Recommendations

Based on OpenClaw's architecture and community best practices:

1. Add Skill Registry Integration

• Integrate with ClawHub API for real-time skill metadata
• Auto-fetch version history, update frequency, download counts
• Community signals: stars, issues, last commit date

2. Implement Automated Security Scanning

• Integrate with clawsec (ClawHub security scanner)
• Auto-scan scripts/ directory for vulnerabilities
• Check for hardcoded secrets, eval/exec patterns

3. Add Utility Benchmarking

• Run comparative tests: baseline vs. with-skill
• Measure task completion time, token efficiency, success rate
• Generate utility scores similar to SkillTester framework

4. Create Skill Dependency Graph

• Map skill dependencies (some skills require others)
• Detect circular dependencies
• Recommend optimal skill combinations

5. Implement Skill Conflict Detection

• Detect skills with conflicting tool usage
• Identify resource contention (browser, file locks)
• Suggest skill compatibility matrix

6. Add Performance Profiling

• Track skill execution time over sessions
• Monitor API usage and costs
• Identify bottlenecks in skill chains

7. Create Skill Reputation System

• Track skill reliability across users
• Aggregate success/failure rates
• Community-rated skill quality scores

8. Implement A/B Testing Framework

• Compare two skills for same task
• Measure which completes faster/better
• Data-driven skill selection

9. Add Skill Update Notifications

• Monitor skill updates in ClawHub
• Alert on breaking changes
• Suggest upgrade timing

10. Create Skill Usage Analytics

• Track which skills are used most frequently
• Identify skill chains and workflows
• Optimize skill loading order

Notes

• This skill adapts EEAT from content/code evaluation to AI agent skill vetting
• Security is elevated to critical importance due to system-level access
• Utility is evaluated comparatively (skill vs. baseline), not absolutely
• Gating mechanisms ensure skills only activate when dependencies are met
• Community signals (downloads, stars, issues) inform Authority dimension
• Prompt quality in SKILL.md directly impacts skill behavior
• Skills with executable scripts require deeper security audit
• This framework is designed for OpenClaw's modular skill architecture