Llm Council

# LLM Council

One AI gives you one answer. That answer feels smart because it was shaped by how you asked. Ask the same question with different framing and you get a different answer, often opposite, equally confident.

The council breaks that loop. Five advisors with different thinking styles answer your question independently. They peer-review each other anonymously. A chairman synthesizes everything into a verdict with a clear recommendation and one concrete next step.

This skill runs the whole protocol inside a single Claude Code session.

## When NOT to convene

If the question fits one of these, just answer directly instead of convening:

- Factual lookup (one correct answer exists).
- Creation task (write, summarize, translate, refactor).
- Trivial choice (e.g. "should I use markdown or plaintext for this note").
- The user has already decided and wants validation. Warn them the council may dissent, then proceed only if they confirm.

If you are unsure, ask the user once: "Is this a real tradeoff you want pressure-tested, or do you want a direct answer?"

## The seven steps

This workflow is rigid. Execute in order. Do not skip, merge, or parallelize across steps.

### Step 1. Scan workspace for context (30 seconds max)

Before framing the question, surface any workspace context the advisors will need. Use Glob and Read. Do not spend more than 30 seconds on this step.

Look for:

- `CLAUDE.md`, `AGENTS.md`, `GEMINI.md` (user or project instructions).
- `memory/**/*.md` (user profile, past decisions, voice).
- Files the user referenced by name or @-mention.
- Recent `council/**/council-report.html` artifacts so you avoid re-counciling ground already covered.

Pick at most 2 to 3 files that would move advisors from generic to grounded. Skip entirely if nothing relevant is present. Never read the whole workspace indiscriminately.

### Step 2. Frame the question

Restate the user's raw question as a single neutral prompt with four parts:

1. Core decision, stripped of emotional lean.
2. User-provided context: facts, numbers, constraints.
3. Workspace context: the 2 to 3 facts from Step 1 that matter here.
4. Stakes: why being wrong is expensive.

Do not inject your own opinion. Do load enough context that advisors can be specific rather than generic.

If the question is too vague to frame ("council this: my business"), ask exactly ONE clarifying question and wait. Then frame.

Save the framed question. Every advisor and every reviewer receives this exact text.

### Step 3. Convene the council (5 advisors in parallel)

Load `references/advisor-prompts.md`. It contains five identity blocks (Contrarian, First Principles, Expansionist, Outsider, Executor) plus a shared wrapper.

Emit all five `Agent` tool calls in a SINGLE assistant message. This is mandatory. Sequential calls defeat the entire point, because later advisors would see earlier ones through context bleed.

For each advisor, the tool call is:

```
Agent(
  description: "<Advisor name>, council r