d0nghyun/ops-retrospect

Retrospect

Review decision paths after task completion. Evaluate whether the chosen approach was optimal and extract reusable lessons.

When to Activate

• After completing a Moderate or Complex task
• User runs /ops-retrospect
• Supervisor detects a task where outcome diverged from expectation

Skip for Trivial tasks (conversation, status checks, simple lookups).

Steps

Step 1: Reconstruct the Path

Identify the decision points from the just-completed task:

| Element | Question | |---------|----------| | Intent | What did the user actually want? | | Complexity | How did I classify it? (Trivial/Moderate/Complex) | | Approach | Direct / Delegate / Collaborate — which did I pick? | | Components | Which agents, skills, teams did I use? | | Branching | Where did I choose between alternatives? |

Step 2: Enumerate Alternatives

For each branching point, list what else could have been done:

Branch: {decision point}
  Chosen:      {what was done}
  Alternative: {option A} — {why it might have been better or worse}
  Alternative: {option B} — {why it might have been better or worse}

Focus on branches where:

• The outcome was suboptimal (errors, revisions, wasted effort)
• The user corrected course mid-task
• Multiple valid approaches existed

Step 3: Evaluate

Score each branching decision:

| Verdict | Meaning | |---------|---------| | Correct | Chosen path was optimal or near-optimal | | Acceptable | Worked but a better path existed | | Suboptimal | A clearly better alternative was available |

Only flag Acceptable and Suboptimal — correct decisions need no action.

Step 4: Extract Lesson

For each non-correct verdict, write a lesson:

Signal:  {what I should have noticed}
Rule:    {decision heuristic for next time}
Scope:   {when this rule applies}

Step 5: Write to Memory

Append to today's memo (vault/memory/YYYY-MM-DD.md):

## Retrospective

### {Task summary}
- **Path taken**: {approach chosen}
- **Verdict**: {Acceptable/Suboptimal}
- **Better alternative**: {what should have been done}
- **Lesson**: {signal} → {rule} (scope: {scope})

If today's memo doesn't exist, create it with only the Retrospective section.

Step 6: Check for Patterns

Read recent memos (vault/memory/) for recurring lessons:

• Same lesson appearing 3+ times → candidate for CLAUDE.md rule update
• Propose the rule change but do NOT auto-modify CLAUDE.md

Output

retrospect_result:
  task: {summary}
  branches_reviewed: {count}
  verdicts:
    correct: {count}
    acceptable: {count}
    suboptimal: {count}
  lessons: [{signal, rule, scope}]
  pattern_detected: true | false

Guardrails

• NEVER auto-modify CLAUDE.md or RULES.md — only propose changes
• NEVER run on Trivial tasks — the overhead exceeds the value
• ALWAYS write lessons to vault/memory, not inline
• ALWAYS be specific — "should have delegated" is too vague; say why and what signal was missed