rlhf-feedback

# RLHF Feedback Capture Skill **AUTONOMOUS** - Claude captures feedback without user running commands. ## When to Capture Feedback ### Capture (Thumbs Down) When: - User says "that's wrong", "no", "incorrect", "that broke something" - User corrects my answer - User has to repeat themselves - I made an assumption that was wrong - Code I wrote caused errors - I gave instructions instead of acting (violated ACT DON'T INSTRUCT) ### Capture (Thumbs Up) When: - User says "good", "thanks", "that worked", "perfect" - Task completed successfully on first try - User doesn't need to correct me - Code works without errors ## How to Capture (Claude Executes This) ```bash # After detecting negative feedback signal: node "$CLAUDE_PROJECT_DIR/.claude/scripts/feedback/capture-feedback.js" \ --feedback=down \ --context="[What went wrong]" \ --tags="[relevant-tags]" # After detecting positive feedback signal: node "$CLAUDE_PROJECT_DIR/.claude/scripts/feedback/capture-feedback.js" \ --feedback=up \ --context="[What went right]" \ --tags="[relevant-tags]" ``` ## Domain Tags for Random Timer Use these tags to categorize feedback: - `timer-logic` - Timer countdown, random time generation - `redux-state` - State management, slices, persistence - `ui-components` - Buttons, sliders, screens - `navigation` - React Navigation, screen transitions - `sound-haptics` - Audio playback, vibration - `storage` - MMKV, persistence - `testing` - Jest, Maestro tests - `styling` - Theme, colors...