Retrospective

The learning loop that prevents the goldfish brain. Run after significant sessions. Takes 5-10 minutes. Pays off every future session.

Step 1: Session Diff

# Node (preferred — zero Python dependency):
node skills/retrospective/scripts/extract-patterns.js

# Python fallback (if Node 18+ unavailable):
python3 skills/retrospective/scripts/extract-patterns.py

Answer:

• What was planned? (from .ohc/plans/)
• What actually happened? (from session log)
• What took longer than expected, and why?
• What was skipped, and why?
• What was added that wasn't planned?

Step 2: Extract Patterns

For each thing that went differently than planned: "Is this a one-off, or would this happen again in a similar situation?"

If it would happen again → capture it.

Good patterns to capture:

• A debugging technique that worked well
• A library API that was confusing (with correct usage)
• A project-specific quirk (e.g., "auth service returns both 401 and 403 for expired tokens")
• A workflow improvement ("next time, check X before starting Y")

Step 3: Update Memory Files

Update .ohc/PROJECT.md if:

• A project gotcha was discovered
• The architecture understanding changed
• A new convention was identified

Format: - [YYYY-MM-DD] {description} → WORKAROUND: {solution}

Update ~/.ohc/USER.md if:

• A personal workflow preference emerged
• A strength or weakness pattern appeared

Step 4: Write Draft Skills

For each novel reusable pattern:

1. Write to .ohc/skills/{kebab-name}.md
2. Keep under 100 lines — just the pattern and when to use it
3. Stays in .ohc/skills/ until promoted:

   ohc skill promote {name}

Step 5: Cursor Users — Generate Rules

If session revealed Cursor-specific learnings: → Use /Generate Cursor Rules to create permanent .mdc rules

Step 6: Summarize

Write 3-sentence summary to .ohc/state/sessions/{id}/summary.md:

1. What was accomplished
2. What was learned
3. Recommended next action

See references/retro-template.md for structured format.