Brainstorm
Decision: $ARGUMENTS
Agent Strategy
Every brainstorm uses 3 core agents in parallel, plus an optional specialist.
Core Agents (always spawn)
| Agent | Role | Mandate |
|---|---|---|
| Explorer | Codebase & Context Analyst | Maps relevant code, existing patterns, constraints, prior decisions |
| Analyst | Trade-off Evaluator | Scores options against quality attributes, produces trade-off matrix + WWHTBT |
| Critic | Devil's Advocate | Assumes the "obvious" choice fails; runs pre-mortem; surfaces risks and biases |
Optional Agent (spawn when relevant)
| Agent | When to spawn | Mandate |
|---|---|---|
| Scout | Unfamiliar tech, external solutions, or user says "what else?" | Researches alternatives nobody mentioned; finds external precedent |
References: agent-roles
Execution
1. Frame the Decision
Before spawning agents, restate the decision clearly and identify:
- Decision statement: What exactly are we deciding?
- Options on the table: 2-4 obvious approaches (agents may add more)
- Priority quality attributes: Performance, security, maintainability, DX, operability, cost, scalability, availability
- Constraints: Timeline, team size, existing tech, compliance, budget
2. Spawn Agents in Parallel
Use Task tool — subagent_type=Explore for Explorer, subagent_type=general-purpose for Analyst/Critic/Scout. Give each their specific mandate and output format from reference files. Spawn all agents in ONE parallel call.
References: workflow
3. Synthesize
After all agents return, combine findings:
- Cross-reference Explorer facts against Analyst scores
- Check if Critic's pre-mortem reveals risks Analyst missed
- Incorporate Scout alternatives if spawned
- Resolve contradictions explicitly — state disagreements and what drives them
References: output-templates, decision-frameworks
Output
Present findings in this order:
- TL;DR — 2-3 sentence recommendation with confidence level
- Trade-off matrix — options scored against weighted quality attributes
- WWHTBT — what would have to be true for each option to succeed
- Pre-mortem & risks — Critic findings, risk register, bias check
- Recommendation — detailed reasoning + ADR-ready summary
References: output-templates
Anti-Patterns
- Don't skip the framing step — vague questions produce vague analysis
- Don't let agents duplicate work — Explorer inspects code, Analyst evaluates, Critic attacks; clear boundaries
- Don't omit the devil's advocate — false consensus is the #1 risk in agent-assisted brainstorming
- Don't present raw agent outputs — the synthesis is the value
- Don't anchor on the first option — always evaluate at least 3 approaches
References: cognitive-biases