Skills are blind after you write them
Static markdown files with no feedback loop. You never know if they're working until something goes wrong.
Activation rate mystery
Skill activation rate can drop to ~20% if descriptions are poor — no data to know it's happening.
Untracked corrections
Users constantly add follow-ups and corrections, but patterns go unnoticed and skills never improve.
Model update drift
When Claude updates, skill outputs quietly degrade. You find out after receiving bad code for weeks.
Token bloat
Skills grow with new features, increasing token waste. No baseline to compare current vs original.
No version science
Comparing two skill versions is guesswork. No way to know which performs better with real data.
Skill Evolver solves all of this
Auto-tracking, reaction detection, health alerts, version history, and A/B testing — zero config.
Everything you need to evolve your skills
From automatic usage tracking to A/B testing — all local, all private.
Skill Usage Tracking
/skill-stats
Auto-collect data every time a skill is invoked via Claude Code hooks — trigger type, tokens, tool calls, duration, files, model version, and skill hash.
Reaction Detection
regex-based
Analyzes the next user message to classify satisfaction. No LLM — no circular dependency.
Health Monitor
/skill-health
Detects degradation signals before they become problems.
Correction Log
/skill-corrections
TF-IDF keyword clustering on correction messages. See patterns humans left behind.
Version Tracking & Rollback
/skill-history · /skill-rollback
Every SKILL.md change is snapshotted with SHA-256 hash. Satisfaction rates linked to each version. Roll back to any previous version in one command.
Skill Guards
ConfigChange hook
Constraints that run on every SKILL.md edit. Warn before drift becomes permanent.
A/B Testing
/skill-ab
Compare two skill versions with real usage data. 50/50 random assignment. Results after 20 runs.
Privacy-First & Zero Config
local SQLite · no telemetry
All data stays in a local SQLite database. Install the plugin and tracking starts automatically — nothing to configure.
How It Works
Three steps. Zero friction. Automatic from day one.
Install the Plugin
Add the repo from marketplace, then install. Two commands — hooks register automatically.
Use Your Skills Normally
Every skill invocation is tracked. Reactions detected. Versions snapshotted. All automatic.
Run Analytics Commands
Check stats, health, corrections, and version history anytime with slash commands.
Slash Commands
Run any of these directly inside Claude Code.
Overview of all skills — top usage, satisfaction rates, token costs, and trend alerts.
Detailed analytics for a single skill: invocations, reactions breakdown, cost, trends.
Raw correction messages + TF-IDF keyword clusters. Spot patterns humans left behind.
Health check all skills. Satisfaction drops, token creep, high cancel rates, model changes.
Detailed health check for one skill with possible causes and rule-based suggestions.
Version timeline with metrics per version. See exactly which change caused satisfaction to drop.
List versions with metrics, pick one, and restore SKILL.md content instantly.
Side-by-side diff of two versions with metrics comparison.
Start an A/B test between current and a new version. 50/50 random assignment per run.
View all running A/B tests with progress toward minimum sample size.
Compare A vs B: satisfaction, tokens, correction rate, follow-up rate, duration.
Export full run data and reactions to CSV or JSON for external analysis.
Built on principles, not hype
Data-driven, not AI-driven
The tool only measures and displays. No LLM used to fix or analyze your skills — avoids circular dependency and AI bias.
Human-in-the-loop always
Data reveals patterns. Humans decide what to change. The tool informs, never auto-edits your skills.
Zero config from day one
Install the plugin. Tracking starts automatically via Claude Code hooks. No setup, no configuration files.
Privacy-first, always local
All data lives in a local SQLite database. Nothing is sent to external servers. Your skill data is yours.
Minimal overhead
Hooks target <100ms execution. SQLite reads stay fast. Your coding workflow is never interrupted.
Like Google Analytics, but for prompts
GA shows bounce rates, drop-off points. Skill Evolver shows satisfaction drops, correction patterns. You decide what to fix.
Built with
Start evolving your skills
Free, open source, private. Install in 60 seconds and get automatic analytics from the first skill run.