skill-governance-and-automation

Skill Governance and Automation

Problem

As a skill repository grows (30+ skills), quality drifts:

• Descriptions become vague (poor activation triggers)
• Evidence sections go stale (cites old competition results)
• Cross-references break ([[links]] to renamed/deleted skills)
• Skills overlap or contradict each other
• No standard summary for quick inventory

The 6-Dimension Skill Audit

For each skill, check:

Dimension 1: Activation Trigger Quality

□ Description starts with "Use when: (1)..." (specific triggers)
□ At least 2 concrete trigger conditions
□ At least 1 "Do NOT use for" exclusion
□ Trigger conditions are testable (not subjective)

BAD: "Tips for Kaggle competitions"
GOOD: "Use when: (1) OOF improved but LB didn't, (2) gap > 1%"

Dimension 2: Evidence Currency

□ Evidence section cites real numbers (not "usually helps")
□ At least 1 competition/experiment as source
□ Evidence date < 6 months old (or marked as "classic, still valid")
□ If competition-specific: competition still active or lesson still applies

STALE: "Validated on S6E4 (2026-05)" → check if still true
CURRENT: "Validated on 20+ competitions through 2026-07"

Dimension 3: Solution Actionability

□ Solution has concrete steps (not just principles)
□ Code examples or commands provided where relevant
□ Decision trees or flowcharts for complex decisions
□ Anti-patterns section (what NOT to do + why)

VAGUE: "Be careful with ensembles"
ACTIONABLE: "If model correlation > 0.97, remove the weaker model"

Dimension 4: Cross-Reference Integrity

□ All [[skill-name]] links point to existing skills
□ No circular dependencies (A → B → A)
□ Related skills are actually related (not just same domain)
□ No two skills cover the same scope without differentiation

Dimension 5: Scope Boundaries

□ Clear scope: what this skill covers
□ Clear exclusions: what this skill does NOT cover
□ No overlap with another skill > 30%
□ If overlap exists: one is the "primary", other links to it