Council

Four-voice structured disagreement for ambiguous decisions and go/no-go calls.

Usage

/council <decision question>
/council <decision question> --round2   # follow-up round, prior verdict optional context

Behavior

1. Extract the real question

State explicitly: what is being decided, what constraints matter, what counts as success. If the question is vague, ask one clarifying question before proceeding.

2. Gather minimal context

Collect only the snippets, metrics, or constraints needed to judge the decision. Skip repo detail for strategic/general questions unless it changes the answer.

3. Form an initial position

Before consulting other voices, write down a starting position, its three strongest reasons, and its main risk. This prevents the synthesis from simply mirroring whichever voice speaks last.

4. Launch three independent voices

Spawn each as a fresh subagent with only the question and the compact context from step 2 — never the full conversation history. This is the anti-anchoring mechanism: a voice that inherits the discussion tends to agree with it.

Apply tiered model delegation: Skeptic and Pragmatist can run on a cheaper model (Haiku-class); reserve the stronger model for Critic or for synthesis if the decision is high-stakes.

Prompt shape per voice:

You are the [ROLE] on a four-voice decision council.

Question: [decision question]
Context: [only the relevant snippets or constraints]

Respond with:
1. Position - 1-2 sentences
2. Reasoning - 3 concise bullets
3. Risk - biggest risk in your recommendation
4. Surprise - one thing the other voices may miss

Be direct, no hedging, under 300 words.

Role emphasis:

5. Synthesize with bias guardrails

• Never dismiss an external view without saying why.
• If a voice changed the recommendation, say so explicitly.
• Always surface the strongest dissent, even when rejecting it.
• Two voices aligning against the initial position is a real signal, not noise.

6. Present the verdict

## Council: [decision title]

**Architect:** [position] - [why]
**Skeptic:** [position] - [why]
**Pragmatist:** [position] - [why]
**Critic:** [position] - [why]

### Verdict
- Consensus: [where they align]
- Strongest dissent: [most important disagreement]
- Premise check: [did the Skeptic challenge the question itself?]
- Recommendation: [synthesized path]

7. Record the outcome

If the verdict is worth keeping, offer /decision-log to capture it as an ADR. Do not write ad-hoc notes elsewhere.

Examples

Input: /council monorepo vs polyrepo for the new services Output: four short positions plus a verdict block recommending monorepo with a named dissent from Critic on CI blast-radius, and an offer to log the decision.

Input: /council ship now or hold for polish Output: Pragmatist and Skeptic align on shipping a scoped subset; verdict flags this as the deciding signal over the initial hold-for-polish position.

Token Optimization

Expected range: 800-2,500 tokens initial (three subagent calls plus synthesis); 150-350 tokens on a --round2 follow-up that reuses the prior verdict instead of re-gathering context.

Caching: none — each council convenes fresh subagents by design, so caching a prior verdict would undermine anti-anchoring. .claude/cache/council/ is not used.

Early exit: if the question has one obvious answer (no real tradeoff), say so in one line and skip convening the council.

Patterns used: tiered model delegation (cheaper model for Skeptic and Pragmatist), progressive disclosure (verdict first, full transcripts only if asked), early exit on non-ambiguous questions.

Edge Cases

• Question is really a code review or implementation task: redirect to /review or standard implementation instead of convening the council.
• User wants unanimous agreement: explain that legible disagreement, not consensus, is the point; report dissent even if all four converge.
• Follow-up round requested: keep the new question narrow and pass only the prior verdict, not the full transcript, to preserve Skeptic value.

Safety

Read-only advisory output. No file writes beyond an explicit, user-approved /decision-log entry. Never feeds subagents the full conversation history.

Related

/brainstorm for open-ended exploration before a decision exists, /first-principles to strip assumptions before framing the question, /decision-log to record the verdict as a durable ADR.