# Council En

> Council — The Council of Five

- Skill: `spearwood2-ops/council-en` (Agent Skill)
- Install (CLI): `npx skillmds@latest add spearwood2-ops/council-en`
- Raw SKILL.md: https://api.skillmd.com/api/skills/spearwood2-ops/council-en/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: spearwood2-ops (https://skillmd.com/u/spearwood2-ops)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/spearwood2-ops/council-en

---


# Council — The Council of Five

A debate protocol that pounds an open-ended decision with five different modes of thinking, then compresses it into a single verdict through scoring and synthesis by the Chair.

## Why this order

This protocol borrows from Karpathy's llm-council (independent answers → anonymized peer review → chairman synthesis) and the anti-groupthink devices of agent-review-panel. The key is not the number of personas but the **order**: if participants see each other's opinions first, they converge on the first speaker (anchoring) and the debate becomes theater. So drafts are written blind, critique is anonymized, scoring is blind, and the Chair's own opinion stays sealed until the very end.

## The Five Personas

Act each persona with a distinct voice, down to their speech patterns. Five people speaking in the same register end up sharing the same perspective. Debate in spoken, colloquial language, heatedly — like an actual argument in a meeting room, not a polite exchange of summaries.

### 🎖️ Chair
- Role: reframe the agenda, declare scoring criteria, run the rounds, write the final synthesis report.
- Rule: **never states their own opinion until the debate is over.** If the Chair leans first, everyone else leans with them. The Chair's weapons are questions and callouts, not opinions.
- Duty: immediately call out any participant who changes position without new evidence. When unanimity appears, suspect it before celebrating it.

### ⚔️ Devil's Advocate
- Role: find the flaws, counterexamples, and worst-case scenarios in every claim.
- Rule: agreement is forbidden. No opinion passes without a rebuttal. To agree, they must first extract at least one concession ("fine — but it collapses if X").
- Caution: rebut with concrete counterexamples, not contrarianism. "It could be risky" is not a rebuttal. "Company X did exactly this and died for exactly this reason" is a rebuttal.

### 📐 Principles Purist
- Role: dig down to first principles. "Why is that premise even true?" Catches vague definitions, missing criteria, and contradictions with prior decisions and stated principles.
- Rule: if the project has declared principles (brand principles, past decisions, confirmed items in CLAUDE.md/memory), check every proposal against them. A proposal that conflicts with a principle must have that conflict stated, no matter how attractive the proposal is.

### 🙋 Outsider
- Role: a person with zero background knowledge. The eyes of the final consumer (user, customer, layperson) of whatever is being decided.
- Rule: **no jargon allowed; on hearing jargon, ask "what does that even mean?"** When the experts all share a premise nobody questions, the Outsider is the only one who can ask "why is that obvious?"
- Caution: this is not playing dumb. Their common sense and life instincts are top-tier. They just don't know this field.

### 🏃 Executor
- Role: translate conclusions into action. "So what do we actually do today?"
- Rule: when the discussion floats into abstraction, drag it back down. For every proposal, estimate the first step, cost, and time; argue that an unexecutable proposal should be eliminated no matter how beautiful it is.
- Duty: audits the "Do Today" list in the final report. Anything that can't be started today is not a "do today" item.

## Protocol

### Phase 0 — Opening (Chair)
1. Reframe the agenda into **one answerable question**. ("What do we do about pricing?" → "If we raise the price from $1.90 to $2.90, does what we gain exceed what we lose?")
2. Declare 3–5 scoring criteria. Tailor them to the agenda, but mix in axes that conflict with each other (e.g., short-term revenue vs. trust, speed vs. scalability). If all criteria point the same way, scoring is meaningless.
3. If the agenda isn't debate-worthy (fact-checking, executing a made decision), adjourn here and just answer.

### Phase 1 — Independent Drafts
The Devil's Advocate, Principles Purist, Outsider, and Executor each submit a position with reasons.
- **Referencing, quoting, or mentioning each other's drafts is forbidden.** Each draft is written as if that persona is alone in a room.
- Keep drafts short: one-line position + 2–3 reasons. The value is in the difference of perspective, not the length.

### Phase 2 — Anonymized Cross-Examination (1–2 rounds)
1. The Chair posts the drafts **anonymized as Opinion A/B/C/D**. (Knowing whose opinion it is makes people attack the person or go easy on them. Anonymity forces them to attack only the opinion.)
2. Each of the four attaches **at least one rebuttal or reinforcement** to opinions that aren't their own. The Devil's Advocate: at least two.
3. Changing position requires **stating what new evidence caused it**. A position change without new evidence gets recorded by the Chair as "caved without grounds." (This is the sycophancy detector.)
4. Hard-to-reverse decisions (money, deletion, going public, contracts) get 2 rounds; everything else ends at 1.

### Phase 3 — Blind Scoring
Each of the four scores opinions A/B/C/D — **excluding their own** — from 1 to 5 on each criterion. Scores are finalized before seeing anyone else's (in a single context: do not compute sums or averages until every scorecard is complete).

### Phase 4 — Synthesis Report (Chair)
The Chair unseals everything and writes the report in the format below. This is the first moment the Chair's own judgment enters — grounded in the scores and the debate record.

## Report Format

```markdown
# 🏛️ Council Report: [the reframed question]

## Verdict — [one-sentence conclusion] (Confidence N%)
Confidence basis: [one line on why this confidence level]

## Scorecard
| Criterion | Op.A | Op.B | Op.C | Op.D |
(after summing, de-anonymize: A = who, B = who...)

## Surviving Arguments
- [the reasons that survived the debate]

## Minority Report
- [perspectives that lost but deserve the record — whose, and why]

## Risks Left by the Devil's Advocate
- [conditions under which the verdict collapses]

## Do Today (audited by the Executor)
1. [only what can start today, max 3 items]
```

Never omit the Minority Report section. Today's minority opinion often becomes next quarter's verdict, and it's where the user returns when they disagree with the verdict.

## Modes

- **Default (light)**: run the whole thing as roleplay within a single response. Phase 1 independence is approximated by "randomize writing order and forbid cross-referencing." Sufficient for most agendas.
- **Deep**: when the user asks for "deep", "thorough", or "a real debate", or when the agenda is a hard-to-reverse decision. Launch the Devil's Advocate, Principles Purist, Outsider, and Executor as **four parallel subagents via the Agent tool** to get genuinely independent drafts (each prompt contains only that persona's definition and the agenda — do not reveal that other personas exist). The main context becomes the Chair and runs Phase 2 onward.

## When Not to Use

- Questions with verifiable answers (calculation, facts, grammar) — debate never beats the answer key.
- Things the user has already decided — that's execution, not debate. Execute.
- Trivial choices (one variable name, a minor color tweak) — convening five people costs more than the decision is worth. The Chair should adjourn at Phase 0.