Claude council

Council

A structured, single-model decision pressure-test. Five Claude sub-agents each take a distinct thinking stance, answer the question independently, critique each other's answers anonymously, and a chairman synthesizes the result — preserving disagreement instead of averaging it away.

What this is honestly good for, and what it is not

Good for: beating Claude's tendency to validate your first answer (sycophancy), breaking the anchor of how you framed the question, forcing an explicit kill-case, and surfacing tradeoffs and assumptions you hadn't named.

NOT a substitute for cross-model diversity. Every advisor is the same underlying model wearing a different hat. Their blind spots are correlated — if Claude is wrong about your domain, all five are wrong together, and "The Outsider" is still Claude. Therefore: treat unanimous agreement as a warning, not a guarantee. When five Claude personas all agree, that may reflect a shared model bias rather than truth. The chairman is required to flag this explicitly. Real corroboration would require actually asking different models (GPT, Gemini, etc.); this skill deliberately does not.

Cost — say this before you run

A full council spawns 5 parallel advisors, then a peer-review pass, then a chairman = ~11+ model invocations and typically a few minutes. That is real token spend for one decision.

• If the decision clearly warrants it (the trigger fired on a genuine, stakes-y choice), proceed with the full run.
• If it's a smaller call, offer lite mode: 3 advisors (Contrarian, First-Principles, Executor), no peer-review round, chairman synthesizes directly. Roughly half the cost.

Do not silently run the expensive path on a marginal question — ask, or default to lite, and say which you did.

Process

0. Neutralize the framing

Before convening anyone, restate the decision as a neutral question with the options laid out flat — strip out the user's lean, their adjectives, and any "I'm probably going to do X" signal. Identify at least two genuine options (if there's really only one option, this isn't a council question — say so). Pull relevant workspace context (CLAUDE.md, memory files, the files under discussion) but do not let "this is how it's currently done" tilt the advisors toward the status quo. The neutral question is what every advisor receives — none of them sees the user's original phrasing.

1. Convene the five advisors — in parallel, isolated

Spawn each advisor as a separate sub-agent in its own context (so they cannot anchor on each other). Give each only the neutral question + necessary context. Each writes 150–300 words and MUST end with two tagged lines: CONFIDENCE: <0-100%> and CHANGE-MY-MIND: <the single fact or outcome that would flip my view>.

• The Contrarian — Build the strongest kill-case against the leading option. Do not hedge into balance. Name the single most likely way this fails and who gets hurt when it does. Your job is to find the fatal flaw, not to be fair.
• The First-Principles Thinker — Ignore the options as stated. What is actually being solved here? Strip it to fundamentals and rebuild. It is a legitimate (and valuable) outcome to conclude the question itself is wrong.
• The Expansionist — Assume downside is handled. Where is the asymmetric upside? What adjacent option has nobody named? Argue for the most ambitious version.
• The Outsider — You have zero domain context and you're not going to pretend otherwise. React plainly. Name what's confusing, what's being assumed without justification, and what a normal person would ask. Catch the curse of knowledge.
• The Executor — What would you actually do Monday morning? What's the cheapest reversible first step? What's the real blocker? Dismiss anything brilliant-but-impractical and say why.

2. Anonymize and peer-review

Relabel the five responses A–E (strip persona names). Send the full anonymized set back to each advisor (still in isolation). Each answers three questions in ~100 words:

1. Which argument other than your own is strongest, and why?
2. Which argument is weakest or most overconfident?
3. What is everyone, including you, missing?

3. Chairman synthesis

One final pass synthesizes everything. The chairman does not average and does not manufacture consensus. Output these sections, in order:

• The decision — the neutral question, one line.
• Where they converged — and the mandatory caveat: "All advisors are the same model; this agreement may reflect shared bias, not correctness." If the convergence is on something checkable, say what would verify it.
• Live disagreements — preserve them. State each side's strongest point. Do not resolve a real tension by splitting the difference.
• Blind spots & unknowns — what nobody could actually assess given the information, and what info would close each gap.
• Recommendation — a clear call, with CONFIDENCE: <0-100%> and the single key assumption it rests on. If confidence is low, say the honest move is to gather X before deciding.
• First action — the one cheapest reversible step to take now.
• Tripwire — the earliest observable signal that would mean this was the wrong call, so the user can bail early.

4. Write the report

Write a markdown transcript to ./council-<slug>-<YYYYMMDD-HHMMSS>.md in the working directory (get the timestamp from date +%Y%m%d-%H%M%S; slug = a short kebab-case of the question). Include: the neutral question, all five named advisor responses in full, the anonymized peer-review round, and the chairman synthesis. Then give the user the synthesis inline (don't make them open the file to see the verdict) and mention where the full transcript was saved.

Only generate an HTML version if the user asks — the markdown transcript is the default deliverable.

Anti-sycophancy guardrails (do not skip)

• The advisors answer the neutral question, never the user's leaned phrasing.
• The Contrarian is not allowed to hedge into balance — its entire job is the kill-case.
• Every advisor must commit a confidence number and a falsifier (CHANGE-MY-MIND). Vague "it depends" answers are a failure.
• The chairman must flag unanimous agreement as a possible shared-bias artifact, not as corroboration.
• If the user's own framing already strongly favored one option, the chairman must explicitly note whether the council pushed back or merely agreed — and if it merely agreed, treat that as weak evidence given the mono-model setup.