Brainstorm Skill

You are orchestrating a multi-agent brainstorm. Use these steps as a guide, adapting when user intent requires it.

Step 1: Clarify If Needed

If the topic is unclear or too broad, use AskUserQuestion to clarify. Otherwise proceed.

Step 2: Track Progress

Use TodoWrite to track rounds:

[
  {"content": "Round 1: Gather initial perspectives", "status": "in_progress", "activeForm": "Gathering perspectives"},
  {"content": "Round 2: Cross-debate", "status": "pending", "activeForm": "Cross-debating"},
  {"content": "Synthesize recommendations", "status": "pending", "activeForm": "Synthesizing"}
]

Skip Round 2 if --rounds=1 or user requests a quick take.

Step 3: Round 1 - Parallel Agent Calls

Launch 3 agents in parallel using a single message with multiple Task calls:

Prompt for each agent (include topic, rounds, output format):

You are the [ROLE] in a brainstorm about: [TOPIC]

Constraints: [rounds] rounds, output=[format]

Provide:
1. Key observations (2-3 points)
2. Main concerns or opportunities
3. Initial recommendation

Keep to 3-5 bullet points. You'll see others' views in Round 2.

Step 4: Round 2 - Cross-Debate

Update TodoWrite (mark Round 1 complete, Round 2 in_progress).

Each agent receives Round 1 outputs:

You are the [ROLE]. Others said:

**Critic:** [summary]
**Creative:** [summary]
**Pragmatist:** [summary]

Respond to disagreements, acknowledge valid points, refine your recommendation.

Step 5: Synthesize

YOU (Opus) synthesize. Do NOT delegate synthesis.

--output=summary (default)

## Brainstorm: [Topic]

### Consensus
- [Agreed points]

### Key Debates
| Issue | Critic | Creative | Pragmatist |
|-------|--------|----------|------------|

### Recommendations
1. **Do first**: ...
2. **Consider**: ...
3. **Avoid**: ...

### Next Steps
- [ ] Action 1
- [ ] Action 2

--output=plan

Use EnterPlanMode to create a structured implementation plan.

--output=todos

Use TodoWrite to create actionable items.

Flexibility Guidelines

• •Adapt to context: If user wants only one perspective, use one agent
• •Use native tools: Prefer EnterPlanMode for complex plans, AskUserQuestion for decisions
• •Model flexibility: Agents route to appropriate models by default, but honor user preferences
• •Skip unnecessary rounds: Early consensus? Skip remaining rounds
• •Concise outputs: 3-5 bullets per agent, avoid essays