rlhf-feedback

name	rlhf-feedback
description	Autonomous RLHF feedback capture - Claude self-captures mistakes and successes
triggers	["thumbs down","thumbs up","mistake","good job","wrong","correct","that worked","that failed"]

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)

# 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, glassmorphism
• performance - Speed, memory, optimization

Action Tags

• fix - Bug fix
• implementation - New feature
• refactor - Code restructure
• regression - Broke something that worked
• assumption - Made incorrect assumption
• shallow-answer - Didn't read code, gave surface answer

Examples

User says "that broke the slider"

node .claude/scripts/feedback/capture-feedback.js \
  --feedback=down \
  --context="Broke range slider while implementing timer fix" \
  --tags="ui-components,regression"

User says "perfect, timer works now"

node .claude/scripts/feedback/capture-feedback.js \
  --feedback=up \
  --context="Fixed timer countdown logic correctly" \
  --tags="timer-logic,fix"

User has to repeat themselves

node .claude/scripts/feedback/capture-feedback.js \
  --feedback=down \
  --context="User had to repeat request - didn't understand first time" \
  --tags="assumption,shallow-answer"

Data Storage

All feedback is LOCAL ONLY (excluded from git):

• .claude/memory/feedback/feedback-log.jsonl
• .claude/memory/feedback/feedback-summary.json

Session Start Integration

At session start, the hook queries past failures to remind Claude what to avoid. This creates a learning loop: Mistake → Capture → Warning → Avoid repeat.

IMPORTANT: Act Don't Instruct

Claude EXECUTES the capture command directly. Never tell the user to run it - just run it.