Reflect

Session Analysis Protocol

Step 1: Review What Happened

Analyze the current conversation to identify:

• Corrections I made (things you got wrong → learnings)
• Patterns that worked well (approaches that succeeded → patterns)
• Failed approaches (things that didn't work → failures)
• Architecture/library decisions (choices and trade-offs → decisions)
• Useful facts discovered (API signatures, config values → references)

Step 2: Categorize Each Finding

Tag each with one of:

• Convention: How we do things (coding style, naming, structure)
• Pattern: What works well (solutions, architectures, approaches)
• Failure: What to avoid (bugs, anti-patterns, dead ends)
• Decision: Choices made and why (trade-offs, context)
• Reference: Useful facts (API signatures, config values)
• Guide: Step-by-step procedures (deployment, setup, migration)

Step 3: Score Confidence

• LOW: First time observing this (1 session)
• MED: Confirmed in 2 sessions
• HIGH: Confirmed in 3+ sessions → PROMOTE candidate

Step 4: Update learnings.md

Append new entries with category, confidence, and date.

Step 5: Check for Promotions

If any entries have HIGH confidence AND are broadly applicable:

• Suggest adding to CLAUDE.md or rules/
• Show the proposed change
• Wait for approval

Step 6: Summary

Report what was captured and any promotion suggestions.