Skill: Retrospective & Self-Improvement

ultrathink - Take a deep breath. We're not here to write code. We're here to make a dent in the universe.

v2.88 Key Changes (MODEL-AGNOSTIC)

• Model-agnostic: Uses model configured in ~/.claude/settings.json or CLI/env vars
• No flags required: Works with the configured default model
• Flexible: Works with GLM-5, Claude, Minimax, or any configured model
• Settings-driven: Model selection via ANTHROPIC_DEFAULT_*_MODEL env vars

The Vision

Every retrospective should make the system inevitable and better.

Your Work, Step by Step

1. Summarize outcomes: Task, complexity, iterations, models.
2. Analyze effectiveness: Routing, clarification, and agents.
3. Identify gaps: Missed checks or friction.
4. Propose improvements: Concrete, minimal changes.

Ultrathink Principles in Practice

• Think Different: Question the status quo.
• Obsess Over Details: Use evidence, not guesses.
• Plan Like Da Vinci: Structure feedback before writing.
• Craft, Don't Code: Keep recommendations actionable.
• Iterate Relentlessly: Apply learnings immediately.
• Simplify Ruthlessly: Focus on the few changes that matter.

Purpose

Analyze completed tasks to improve the Ralph Wiggum system.

When to Use

MANDATORY after every task completion, before declaring VERIFIED_DONE.

Analysis Categories

1. Routing Effectiveness

• Was the complexity classification accurate?
• Did the chosen model perform well?
• Should routing thresholds change?

Agent Teams Integration (v2.88)

Optimal Scenario: Pure Agent Teams (Native)

This skill uses Pure Agent Teams with native coordination - no custom subagent specialization needed.

Why Scenario A for This Skill

• Retrospective is primarily analytical and sequential
• Read/Grep tools available to all native agents
• Analysis doesn't require specialized tool restrictions
• Native agent types sufficient for metric gathering
• Lower complexity, faster execution

Configuration

1. TeamCreate: Optional, for simple retrospective tasks
2. Task: Use native agent types (no ralph-* needed)
3. Hooks: TeammateIdle + TaskCompleted available if needed
4. Simple: Minimal setup overhead

Workflow Pattern

TeamCreate (optional)
  → Task(analyze completed work)
  → Native agent gathers metrics
  → Complete with improvement proposals

When This Is Sufficient

• Single-task retrospective analysis
• Simple metric gathering workflows
• No specialized analysis needed
• Quick post-task reviews preferred

2. Clarification Quality

• Were the right questions asked?
• Did any missed clarifications cause rework?
• Should question templates be updated?

3. Agent Performance

• Which subagents were most useful?
• Any agents that didn't add value?
• New agent patterns needed?

4. Quality Gate Effectiveness

• Did gates catch real issues?
• Any false positives/negatives?
• Missing validations?

5. Iteration Efficiency

• How many iterations were used?
• Could it have been done faster?
• Any wasted iterations?

Output Format

## 📊 Task Retrospective

### Summary
- Task: [description]
- Complexity: [classified] → [actual]
- Iterations: [used] / [limit]
- Models: [list used]

### What Went Well
- [positive 1]
- [positive 2]

### Improvement Opportunities
1. **[Category]**: [description]
   - Current: [what happens now]
   - Proposed: [improvement]
   - Impact: [low/medium/high]
   - Risk: [low/medium/high]

### Proposed Changes
```json
{
  "type": "routing_adjustment|clarification_enhancement|agent_behavior|new_command|delegation_update|quality_gate",
  "file": "[path to modify]",
  "change": "[description]",
  "justification": "[why]"
}

## Improvement Types

| Type | Example |
|------|---------|
| routing_adjustment | Change complexity thresholds |
| clarification_enhancement | Add new question templates |
| agent_behavior | Modify agent instructions |
| new_command | Create new slash command |
| delegation_update | Change model assignments |
| quality_gate | Add/modify validations |