Run your hard decision past an AI council in Claude Code

---
name: llm-council
description: Runs a five-adviser decision council as parallel subagents on Opus (contrarian, first-principles thinker, expansionist, outsider, executor), has the same five challenge each other anonymously, then gives one chairman's verdict and saves an HTML report. Use when the user types /llm-council, says "council this", or explicitly asks for a council on a high-stakes decision. Not for lookups or quick yes/no questions.
argument-hint: "[the decision you're stuck on]"
model: opus
---

# LLM council

DECISION: $ARGUMENTS

If the decision above is empty or one vague line, ask the user one question at a time (their situation, their constraints, what a good outcome looks like) and wait for each answer. Do not start any agent until you have the decision.

The council is five agents and you. You are the chairman. Never write an adviser's answer or challenge yourself, and never start more than five agents.

Copy this checklist and tick it off as you go:

```
Council progress:
- [ ] 1. Five advisers answered (5 Agent calls, one message)
- [ ] 2. The same five challenged each other (5 SendMessage calls, one message)
- [ ] 3. Chairman's verdict written
- [ ] 4. HTML report saved
```

## The advisers

- **Contrarian**: looks only for what will fail. Every reason this is wrong, what breaks first, the worst plausible outcome.
- **First-principles thinker**: questions the user's assumptions, drops the standard frameworks, reduces the problem to what is true and rebuilds the answer.
- **Expansionist**: finds the upside the user missed. The outcome if this works far better than expected, and what a bigger version opens up.
- **Outsider**: knows nothing about the user's field. Challenges what insiders take for granted and asks the plain questions they stopped asking.
- **Executor**: ignores strategy. Exactly what the user does on Monday morning and through the week: the email, the conversation, the file, the decision to put off.

## 1. The council answers

Send five Agent calls in a single message, one per adviser, so they run in parallel. On every call use `subagent_type: general-purpose`, `model: opus` and `run_in_background: false`, and make the description the adviser's name (for example "Contrarian").

Each prompt contains the decision, that adviser's lens from the list above, and these rules: "You are one adviser on a five-person council. Stay in your lens; do not balance it with other views. Do not name or describe your role anywhere in the answer. Do not use any tools. Answer in under 300 words."

Write down the `agentId` each call returns, next to the adviser's name. You need it in step 2.

## 2. They challenge each other

Run `pwd`, then create a folder `council-YYYYMMDD-HHMM/` inside it (today's date and time). Build every file path from the `pwd` output exactly, character for character.

Shuffle the five answers and label them Response A to E. Keep the mapping to yourself. Delete any line or phrase that reveals an adviser (for example "Contrarian read:" or "As the outsider"), and change nothing else. Write the decision and the five lettered responses to `responses.md` in the folder, once.

If the SendMessage tool isn't loaded, load it with ToolSearch (`select:SendMessage`). Then send five SendMessage calls in a single message, one to each adviser's `agentId` from step 1. Do not start new agents for this step. Each message:

"Round two. Your answer is Response [LETTER]. Read [FULL PATH TO responses.md] with the Read tool and skip your own. Rank the other four from 1 to 4. For each, give its strongest point and its biggest flaw, judged through your own lens. Then name the one point in the council you most disagree with, and why. Use no tool other than Read. Under 300 words."

The replies come back one by one. Wait until all five advisers have replied before step 3.

## 3. Chairman

Read all five answers and all five challenges. First, say which response the council ranked highest overall (add up the ranks, lowest wins) and where it disagreed most.

Then the verdict, under 250 words. No hedging, no "both sides".

- **Decision**: what to do
- **Strongest reason**: the one reason that carries it
- **Biggest risk**: the one thing to watch
- **Next 7 days**: one specific step
- **What would flip it**: the one assumption that, if wrong, changes the decision

## 4. Report

Write `report.html` in the same `council-YYYYMMDD-HHMM/` folder with the Write tool. One self-contained file with no external scripts: the decision, the chairman's verdict at the top, then the five answers and the five challenges in full (named, not summarised), the letter-to-adviser mapping and the rank totals. Readable on a phone.

Then show the chairman's verdict in chat, with the report path.

If any step did not run through the five adviser agents, say so at the top of the verdict and in the report.