Coach Review

Analyze and display pending learning candidates from the coach system.

Steps

1. Read the candidates file:

   cat ~/.claude-coach/candidates.json 2>/dev/null || echo '{"pending":[]}'
   

2. For candidates with needs_llm_analysis: true or candidate_type: "raw_correction":

   These contain raw friction data that needs YOUR analysis. Look at:

   • The raw_data.message - what the user said
   • The raw_data.context_before - what happened before
   • The raw_data.patterns_matched - what patterns were detected

   Analyze and generate a proper proposal:

   • Understand what Claude did wrong
   • Determine the correct behavior
   • Create a specific, actionable rule
   • Set an appropriate confidence score (0.7-0.95)

3. For other candidates: Display as-is with:

   • ID and title
   • Type (rule/checklist/snippet/skill)
   • Confidence score
   • Evidence that triggered it
   • The proposed trigger and action

4. Your recommendation for each: Based on your analysis, indicate:

   • ✅ Recommend approve - clear pattern, actionable rule
   • ⚠️ Needs refinement - good signal but vague proposal
   • ❌ Recommend reject - false positive or too specific

5. Show available actions:

   • /coach:approve <id> - Accept and apply this proposal
   • /coach:edit <id> - Modify before approving
   • /coach:reject <id> <reason> - Reject with reason

Format output as a clear, readable summary. Group by recommendation.