--- name: council description: > Stress-test a decision, plan, or question with a 5-advisor council. Spawns FIVE independent advisor personas (Contrarian, First-Principles, Expansionist, Outsider, Executor) who answer blind to each other, then peer-reviews their answers anonymized and shuffled, then hands back a single verdict: best recommendation, biggest blind spot, what everyone missed, and one concrete next step. Produces a skimmable HTML report plus the full transcript. Use when the user says "council this", "/council", "run the council on X", "get the council's take", "stress-test this decision", "pressure-test this plan", or pastes a question/prompt and wants it debated before they trust the answer. Best for genuinely consequential, multi-variable, or reversible-but-costly decisions — not quick factual lookups. --- # The Claude Council Five advisors debate the question so the answer can be trusted. The point is **structured disagreement** — five genuinely different lenses, peer-reviewed anonymously so no advisor gets special treatment, then synthesized into a verdict that names the strongest path AND what the whole room missed. Core rule: **let disagreement BE the signal — do not synthesize the conflict away.** If the advisors split, the verdict says so and explains the fork. ## When to use it Use for decisions where being wrong is expensive and the answer isn't obvious: strategic bets, big trade-offs, "should I do X or Y", plans about to be committed, framings that feel shaky. Skip it for quick factual questions, mechanical tasks, or anything with one obviously-correct answer — the council is expensive and adds nothing there. If the question is genuinely trivial, say so and answer directly instead of convening. ## Inputs - **Required:** the question, decision, or prompt to council. If the user just said "council this" with no clear target, the target is the immediately preceding topic of the conversation — restate it in one line and proceed (don't interrogate). - **Context is gold.** Pull in whatever's relevant and available: the conversation so far, pasted material, and any relevant project files (a context.md, a brief, recent decisions). Pass this context to every advisor. A council with no context produces generic advice. - **Optional:** a stakes/reversibility note from the user. If absent, infer it. ## Workflow ### Step 1 — Frame the question Write a single crisp **decision statement** (1-2 sentences) plus a short **context block** (what's true, constraints, what's already been tried/decided). This exact framing goes to all five advisors verbatim so they're answering the same question. If the question is ambiguous in a way that changes the answer, pick the most useful interpretation, state it, and proceed. ### Step 2 — Convene the council (5 independent advisors, IN PARALLEL) Spawn **all five advisors in a single message** (five `Agent` tool calls, `subagent_type: general-purpose`) so they run concurrently and **blind to each other**. Never let one advisor's output leak into another's prompt. Give each the *same* decision statement + context block, then its own lens. Each advisor prompt must end with this instruction: > Return your answer as: **(1) Position** — your one-paragraph recommendation. **(2) Reasoning** — the 2-4 load-bearing points. **(3) Sharpest insight** — the single thing your lens sees that others probably won't. **(4) Confidence** — High/Medium/Low and why. Be specific and decisive. Do not hedge into mush. ~250-350 words. You are NOT writing to the user; you are returning raw council input. The five lenses (use these exact roles): 1. **The Contrarian — "What could fail?"** Your job is to find the failure modes, the hidden costs, the ways this goes wrong. Steelman the case *against*. What is everyone too excited to see? What breaks at scale, under stress, or when an assumption is wrong? Be the friend who says the hard thing. 2. **The First-Principles Thinker — "What problem are we really solving?"** Strip the question to its root. Are we solving the actual problem or a proxy for it? What assumptions are baked in that don't have to be? Rebuild the answer from fundamentals, ignoring how it's "normally" done. 3. **The Expansionist — "What upside are we missing?"** Think bigger. Where's the asymmetric, 10x, or compounding version of this? What option are we not even considering because it feels too ambitious? If this works better than expected, what does that unlock? 4. **The Outsider — "Does this make sense to someone new?"** You have zero context-debt and no sunk cost. Where's the jargon hiding a weak idea? What would a smart person hearing this for the first time find confusing, unconvincing, or obviously off? Apply plain common sense the insiders have rationalized away. 5. **The Executor — "What would you actually do Monday morning?"** Ignore theory. What's the concrete first move, the sequence, the owner, the timeline? What's the smallest version that tests the core risk this week? Translate everything into action — and flag where the plan is too vague to execute. If any advisor returns `null` (skipped), note it and proceed with the rest — a 4-advisor council is still valid; a <3 council is not (re-run). ### Step 3 — Anonymized peer review Label the five responses **A–E in shuffled order** (do NOT preserve advisor order — strip the role names so review is blind). Then run the review. Default: spawn **3 reviewer agents in parallel**, each given the full anonymized set and asked the same three questions. (Lightweight alternative when stakes are low: do the review yourself in one pass.) Each reviewer answers: 1. **Which response is strongest, and why?** (cite by letter) 2. **Which response has the biggest blind spot?** (cite by letter + name the blind spot) 3. **What did ALL FIVE miss?** — the question none of them answered. Tally the votes. Where reviewers agree, that's a strong signal. Where they split, that's also signal — keep it. ### Step 4 — Synthesize the verdict You (the main thread) write the verdict. Do not just average. Read the advisor answers and the peer review, then produce: - **🏆 Best recommendation** — the path the evidence + review most support. If two advisors genuinely conflict and both are right under different assumptions, present the **fork** and the condition that decides it. Don't fake consensus. - **🔍 Biggest blind spot** — the most dangerous weakness in the leading recommendation (sharpened from the peer review). - **⚠️ What everyone missed** — the angle no advisor surfaced (from review Q3 + your own read). This is often the most valuable line in the whole report. - **✅ One concrete next step** — a single, specific, do-it-this-week action. Owner + first move. Not a list. Keep the verdict tight and decisive. This is what the user reads first. ### Step 5 — Deliver 1. **In the terminal:** print the verdict (the four blocks above) — concise, skimmable, no preamble. 2. **Write the HTML report** using `references/report-template.html` as the structure: verdict at top, then the five advisor cards, then the peer-review tally, then the full transcript collapsed at the bottom. Self-contained (inline CSS), opens in any browser. 3. **Save outputs.** Save the report as `council-YYYY-MM-DD-.html` plus a companion `.md` transcript, to wherever the project keeps decision records (or the current working directory if there's no obvious home). Tell the user the path and offer to open the HTML. ## Guardrails - **Independence is the whole point.** Advisors must not see each other's work in Step 2. Peer review must be anonymized in Step 3. If you collapse these, you've built an echo chamber, not a council. - **Don't sand down conflict.** A verdict that hides a real split is worse than useless — it launders disagreement as confidence. Name forks explicitly. - **No hedge-mush.** Every advisor and the verdict must commit to a position. "It depends" is only acceptable with the deciding variable named. - **Scale to stakes.** Trivial question → say it doesn't need a council and answer directly. Huge irreversible decision → consider a second review round or more advisors. - **Ground in context.** Generic advice is the failure mode. Feed real context in; if there's none, say the verdict is context-light.