experimentlisted

# Experiment Autonomous metric-driven iteration loop inspired by Karpathy's autoresearch pattern. The core cycle: modify code, measure a metric, keep improvements, discard regressions, repeat. Works with any file, any numeric metric, and any direction (minimize or maximize). Git-based safety guarantees: commit before evaluate, reset on regression -- regressions are never persisted. Produces a structured report of what worked, what failed, and cumulative gains. ## When to Use /experiment vs Other Skills | Goal | Use | |------|-----| | Improve test coverage toward a target | `/cover --target=N%` (specialized coverage loop) | | Fix a specific bug | `/fix-bug` (observation-driven debugging) | | Build a feature end-to-end | `/develop` (gated pipeline) | | Optimize ANY measurable metric | `/experiment` (this skill) | | Understand agentic loop theory | `agent-loops` (pattern reference) | Use `/experiment` when the task has a single numeric metric, a clear direction (minimize or maximize), and benefits from autonomous iteration rather than manual tuning. ## Quick Start ``` # Reduce API latency below 100ms /experiment src/api/handler.ts --metric "npm run bench -- --json | jq '.results[0].mean'" --minimize --goal 100 --unit ms # Increase Lighthouse performance score to 95 /experiment src/pages/ --metric "npx lighthouse http://localhost:3000 --output json | jq '.categories.performance.score * 100'" --maximize --goal 95 # Minimize bundle size over 15 iterations /experiment src/ -