Skill Auditor — Quality Scanner & Curator

"装 skill 之前先审。市面 skill 质量参差，安全扫描是刚需。"

10-Point Quality Scan

#	Check	What It Detects
1	YAML validity	Broken/missing frontmatter
2	Description quality	Vague triggers, missing "when to use"
3	Token efficiency	Bloat >500 lines, redundant instructions
4	Prompt injection	Hidden commands in description or body
5	Dangerous commands	rm -rf, .env writes, --force patterns
6	Data exfiltration	Curl to unknown domains, file reads outside project
7	Obfuscated code	Base64 blobs, zero-width chars, homograph attacks
8	Dependency risk	References to unvetted MCP servers or scripts
9	Trigger accuracy	Does description match actual behavior?
10	Overlap detection	Duplicates existing installed skills?

Scoring

Each check: pass (1) or fail (0). Total: 0-10.

Score	Verdict	Action
9-10	Trusted	Safe to install
7-8	Caution	Install, monitor first 5 runs
5-6	Risky	Read source before installing
0-4	Blocked	Do not install

Skill Recommender

When user asks "which skill for X?":

1. mcp__memory__search_nodes → find similar past tasks
2. Scan installed skills for best match
3. If none → search skills.sh marketplace
4. Return top 3 with audit scores

Curator Mode

Maintains a curated skill index (stored in memory.db via memory-lora):

• Installed skills with audit scores
• Duplicate/deprecated markers
• Recommendations per task type

Before Install Checklist

• Audit score ≥7
• Not duplicate of installed skill
• Dependencies are trusted
• No prompt injection detected
• Description triggers are clear

Trigger

• User runs npx skills add or /plugin install
• User browses marketplace ("which skill for X?")
• User asks "is this skill safe?"
• New skill detected in skills/ directory

Lifecycle

• Pre-install: skill-auditor scans safety (this skill)
• Post-use: skill-judge evaluates effectiveness (complementary partner)
• memory-lora: 🔜 Audit scores stored as baseline for future comparisons