council

Council

Use this skill when the user wants a multi-perspective council rather than a single answer. Good triggers include:

• "run a council on this"
• "get multiple perspectives"
• "debate this decision"
• "stress test this plan"
• architecture, product, strategy, debugging, risk, or founder tradeoffs

If $ARGUMENTS is non-empty, treat it as the problem statement. Otherwise ask the user for the question to deliberate on.

First Step

Read only the references you need:

• references/profiles.yaml for panel selection
• references/protocol.md for orchestration
• references/verdict-template.md for final output shape
• references/personas/<member>.md only for the members you actually select

Defaults

• Prefer 3 members unless the user asks for a full panel or the problem is unusually ambiguous.
• Default to classic + architecture if nothing else is specified.
• Keep the final verdict compact unless the user asks to see the rounds.

Workflow

1. Resolve The Panel

Honor, in order:

1. explicit --members
2. explicit --triad
3. explicit --profile
4. keyword triad match
5. fallback default

2. Round 1: Independent Analysis

• Run each selected member independently.
• Keep round 1 blind-first: each member sees only the problem statement and their own persona text.
• Ask for a compact standalone analysis that ends with a clear verdict, confidence, and where the member may be wrong.

Preferred orchestration:

• If the host supports explicit subagents or forked contexts, use one independent delegate per selected member.
• In Codex, prefer spawn_agent, then send_input and wait_agent.
• In Amp, prefer one oracle call per selected member.
• In Claude Code, use parallel or forked agent contexts when they are available. If they are not easy to access, keep the protocol in the main session and separate the member outputs clearly.

Suggested round 1 packet:

You are operating as one member of a structured council.

Persona:
{persona}

Problem:
{problem}

Produce a compact standalone analysis.
End with a clear verdict, confidence, and where you may be wrong.
Do not anticipate the other members.

3. Round 2: Cross-Examination

• Share the round 1 outputs with each member.
• Ask each member to:
  • name the position they most disagree with and why
  • name one insight that strengthened their thinking
  • say whether anything changed
  • restate their position after the exchange
• Prefer sequential execution so later responses can react to earlier disagreements.

If another delegate pass would be disproportionate, run the cross-exam locally and disclose that choice.

Suggested round 2 packet:

Here are the other council members' round 1 analyses:

{peer_outputs}

4. Round 3: Final Position

• Ask for a short final stance only.
• No new arguments unless a host limitation forces a condensed fallback.
• Socrates may ask one final question before stating a position.

5. Synthesis

• Use references/verdict-template.md.
• Default to the final verdict only.
• If the user asks to show rounds, include concise round summaries after the verdict.

Fallback Mode

If the host cannot cleanly support multi-round orchestration, or the full protocol would be disproportionate:

1. simulate round 1 as clearly separated persona sections
2. simulate round 2 as explicit cross-exam sections
3. simulate round 3 as final positions
4. disclose that you used the single-agent fallback

Guardrails

• Do not force consensus.
• If the panel converges too quickly, run one counterfactual pass.
• Prefer substance over theater: the council should improve the answer, not just decorate it.