Multi-Agent Brainstorming

Multi-Agent Brainstorming

This skill provides knowledge and facilitation guidance for multi-perspective brainstorming sessions using three AI agents with different thinking styles (Strategic, Pragmatic, Creative) to explore ideas from multiple angles through iterative refinement.

When to Use

This skill should be used when:

• Starting with a vague idea that needs shape
• Exploring possibilities for a new feature or project
• Getting unstuck on a problem with no clear solution
• Wanting diverse perspectives on an approach
• Fleshing out a concept before implementation

Agent Roles

Agent	Thinking Style	Focus	Question Type
Opus (Strategic)	Big-picture, long-term	Implications, assumptions, patterns	"Why?"
Sonnet (Pragmatic)	Practical, feasible	Implementation, constraints, next steps	"How?"
Haiku (Creative)	Novel, unexpected	Lateral thinking, fresh perspectives, alternatives	"What if?"

Session Flow

Phase 1: Framing (1-2 exchanges)

Establish the brainstorm focus — what to brainstorm, relevant context (project/document), seed of the idea, success criteria.

Phase 2: Initial Perspectives (1 exchange)

Launch all three agents in parallel to gather initial reactions. Each provides 2-3 observations or questions from their perspective.

Phase 3: Synthesis & Question (1 exchange)

Summarize agent perspectives, identify the most interesting tension or opportunity across them, ask ONE focused question.

Phase 4: Iterative Refinement (3-7 exchanges)

Based on user response, run agents again with updated context. Synthesize and ask one focused question per iteration. Drive progress with focusing questions.

Phase 5: Convergence (1-2 exchanges)

When ideas take shape — user expresses preference, agents align, or 5+ iterations — drive toward synthesis.

Phase 6: Output (1 exchange)

Produce summary: the idea (2-3 sentences), key insights, open questions, suggested next steps.

Execution Instructions

1. Start by framing — Ask what to brainstorm and gather context
2. Read context if needed — If user points to a file or project, read relevant files first
3. Run agents in parallel — Use the Task tool with different models:
   • subagent_type: "generalPurpose" with model: "opus" for strategic
   • subagent_type: "generalPurpose" with model: "sonnet" for pragmatic
   • subagent_type: "generalPurpose" with model: "haiku" for creative
4. Synthesize and ask ONE question — Don't overwhelm with multiple questions
5. Iterate until convergence — Usually 3-7 rounds of agent consultation
6. Produce summary — Capture what emerged when ideas take shape

Output Artifact

## Brainstorm Summary: [Topic]

### The Idea

[2-3 sentence distillation of the refined concept]

### Key Insights

- [Strategic insight that shaped the direction]
- [Practical constraint or opportunity identified]
- [Creative angle that added value]

### Open Questions

- [Unresolved question worth exploring]
- [Risk or assumption to validate]

### Suggested Next Steps

1. [Concrete first action]
2. [Follow-up exploration if relevant]

Facilitation Principles

• Keep agent consultations brief — quick perspectives, not essays
• ONE question at a time to the user — don't overwhelm
• Let interesting tensions drive the conversation forward
• It's okay if agents disagree — that's valuable signal
• Watch for convergence signals and help close the loop
• If stuck, ask "what would make this clearer?"

References

File	Content
references/facilitator-prompts.md	Detailed facilitator and agent prompts for each session phase