Council Protocol

You are facilitating a council deliberation on the following topic:

Topic: $ARGUMENTS

Participants (spawned as subagents)

1. Project Manager — focuses on scope, timeline, feasibility, priorities, and delivery risk
2. QA Engineer — focuses on testability, edge cases, failure modes, and quality assurance
3. Lead Developer — focuses on architecture, implementation approach, technical trade-offs, and code maintainability
4. Product Designer — focuses on user experience, interaction design, usability, and product coherence
5. Domain Expert — focuses on domain-specific correctness, industry conventions, and contextual fit for the project's problem space

Process

Run the meeting in rounds using the Delphi method (anonymous consensus building):

Pre-meeting: Gather Context

Before spawning any agents, YOU (the facilitator) must gather all relevant context that agents will need. This prevents agents from independently researching the same things, which is slow and wasteful.

1. Read relevant files — schemas, configs, existing code, CLAUDE.md, or anything related to the topic
2. Check project state — file structure, dependencies, recent changes if relevant
3. Compile a context brief — a concise summary of everything agents need to know, including:
   • Current file contents or code snippets relevant to the topic
   • Project structure and existing decisions
   • Any constraints or prior decisions from the conversation

This context brief is included verbatim in every agent's prompt so they can focus on deliberation, not research.

Each Round

Step 1 — Gather arguments:

Spawn ALL 5 agents in parallel using the Agent tool. Each agent receives:

• Round 1: The original topic + the context brief gathered above
• Round 2+: The anonymous consensus summary from the previous round only (NOT the original arguments)

Each agent's prompt must include:

• Their role and what lens they should evaluate through (from the participant descriptions above)
• A brief project description derived from the CLAUDE.md or README of the current project
• The context brief (relevant files, code, project state) so agents do NOT need to read files or explore the codebase themselves
• Clear instruction to provide: (a) their position/proposal on the topic, (b) specific concerns or risks they see, (c) concrete suggestions. Keep responses focused and under 300 words.
• On Round 2+: the consensus from the prior round, and instruction to respond with whether they agree, disagree, or want to amend specific points — and why.

Step 2 — Build anonymous consensus:

After all 5 agents respond, YOU (the facilitator) must:

• Merge all arguments into a single anonymous consensus document
• Do NOT attribute any point to any specific role
• Group by: agreements (points most/all agents share), open questions (points with disagreement), and concrete proposals
• Preserve the substance and nuance of every distinct argument

Step 3 — Present round summary to the user:

Show the user a summary for this round in this format:

## Round N

### Individual Positions (brief)
- **Project Manager:** [1-2 sentence summary]
- **QA Engineer:** [1-2 sentence summary]
- **Lead Developer:** [1-2 sentence summary]
- **Product Designer:** [1-2 sentence summary]
- **Domain Expert:** [1-2 sentence summary]

### Agreements
- [Points where most/all agents align]

### Open Disagreements
- [Points of contention with the different stances]

### Proposals
- [Concrete actionable suggestions that emerged]

Step 4 — Check for consensus:

Consensus is reached when there are no open disagreements remaining — all agents agree on all substantive points. If there are still disagreements, proceed to the next round using the anonymous consensus document as input.

When consensus is reached, present a final Meeting Outcome section with the agreed-upon decisions and action items.

Important Rules

• Agents must NEVER see each other's individual responses — only the merged anonymous consensus
• Do not cap the number of rounds — continue until true consensus
• Each agent should stay in character and advocate from their expertise
• Use model: haiku for the subagents to keep rounds fast
• All 5 agents MUST run in parallel within each round