Skill Evolver

Analytics and health monitoring for Claude Code skills.

Like Google Analytics for your skills — tracks usage, detects user reactions, monitors health trends, and enables A/B testing. All data stays local in SQLite.

Install

# install
claude plugin marketplace add https://github.com/haidang1810/skill-evolver
claude plugin install skill-evolver
# update
claude plugin marketplace update skill-evolver
claude plugin uninstall skill-evolver
claude plugin install skill-evolver

Features

Skill Usage Tracking

Automatically tracks every skill invocation via hooks:

• Skill name, trigger type (explicit/auto), arguments
• Token consumption, tool calls, duration
• Skill version hash, model used

User Reaction Detection

Classifies user’s next message after a skill runs:

• Satisfied — “thanks”, “looks good”, topic change
• Correction — “no, actually…”, “change that…”
• Follow-up — “also check…”, “you forgot…”
• Retry — Same skill invoked again immediately
• Cancel — Interrupted mid-execution

Slash Commands

Health Monitoring

Detects skill degradation:

• Satisfaction trend drops (> 15%)
• Token creep (> 30% above baseline)
• High cancel rate (> 10%)
• High correction rate (> 25%)
• Model changes
• SKILL.md file changes

Skill Guards

Validates SKILL.md changes against drift thresholds:

• Line count limit (500 lines max)
• Line drift from baseline (30% max growth)
• Step count drift (±3 steps from baseline)
• Description length (200 chars max)
• Token budget drift

Version Tracking

Automatic SKILL.md snapshots with:

• Content hash + line count
• Parent version linkage
• Metrics per version
• One-command rollback

A/B Testing

Compare two skill versions with real usage data:

• Random 50/50 version assignment
• Satisfaction, tokens, correction rate comparison
• Statistical significance caveat

Design Philosophy

1. Data-driven, not AI-driven — Measures and displays, never uses LLM to auto-fix
2. Human-in-the-loop — You decide what to change based on data
3. Zero config — Install and it tracks automatically
4. Minimal overhead — Async hooks, never blocks Claude
5. Privacy-first — All data in local SQLite, nothing sent externally

Tech Stack

• Runtime: Node.js (ES Modules)
• Storage: SQLite via better-sqlite3
• Keyword extraction: Built-in TF-IDF (no deps)
• Hashing: Node.js crypto

Data Storage

Database: /data/skill-evolver.db

Tables: skill_runs, reactions, skill_versions, ab_tests, ab_runs, guard_configs

License

MIT