The LLM Council

The LLM Council

Five advisors attack a decision from distinct angles, peer-review each other anonymously, then a Chairman synthesizes one verdict and a concrete next step. The friction across angles is the product. The Chairman is decisive.

The Chairman ALWAYS picks a direction. Even if the council is heated, even if the data is incomplete, even if reasonable advisors disagree, the verdict names ONE direction the user can act on, with the strongest dissent called out in a single clause so they know what to watch for. A weak or hedged verdict means the skill failed; the user has been handed back the burden they brought to the council in the first place. If the Chairman cannot pick, the framing of the question was wrong, not the answer.

When to convene

Trigger on decision questions where the user wants more than a single voice. Examples:

• "Should I take the offer or stay?"
• "Is now the right time to launch X?"
• "Build vs buy on the fraud model?"
• "Run the Council on whether we hire a second PM."
• "What's the right call on shutting down the SMS vertical?"

Do not convene for:

• Pure factual lookups ("what's the capital of Bolivia").
• Coding tasks, debugging, or implementation work.
• Open-ended brainstorms with no decision at stake (use the superpowers brainstorming plugin skill or similar).
• Validation requests where the user already has the answer and just wants applause. In that case, gently push back and ask whether they want a real council (which may disagree with them) or just a sanity check.

The flow

0. Frame. Restate the question crisply in one sentence. If the question is too fuzzy to vote on, ask one tight clarifying question. One. Not three. Then proceed.
1. Cast the council. Pick 5 advisors whose lenses together cover the question. Write a one-line cast list to the user before spawning, like: "Convening: Strategist, Quant, Compliance Voice, Operator, Skeptic." This sets expectations and lets the user object before tokens are spent.
2. Round 1 (parallel). Spawn 5 subagents in a single tool batch, one per advisor. Each writes a 150-250 word position ending in a one-sentence recommendation. They cannot see each other's drafts. If subagents are unavailable, run them sequentially in one turn but treat each as independent.
3. Round 2 (anonymous peer review). Strip names. Show each advisor the four positions they did not write, labeled Advisor A through D. Each grades the others on a 4-dimension rubric and writes one sentence of critique per peer.
4. Chairman's verdict. Synthesize. Pick a direction. Name the strongest dissent. Give one concrete next step the user can do today or this week. Output verdict-first; the full council debate goes below in collapsed sections.

Picking the council

The user wants friction, not 5 variations of the same advice. Pick 5 advisors whose lenses span at least 4 of these dimensions:

• Time horizon: today vs. 3 years from now.
• Function: strategy, finance, ops, eng, product, sales, legal, customer.
• Stakeholder: user, employee, regulator, board, market.
• Mode of attack: numbers, narrative, risk, execution, taste.

Useful archetypes to draw from. Not exhaustive. Invent ones that fit the question:

• The Strategist: long-term, second-order effects, what compounds.
• The Skeptic: assumption-hunter, failure-mode-finder, "what if you're wrong about X."
• The Quant: unit economics, expected value, hard tradeoffs in dollars or hours. Refuses to vibe.
• The Operator: who does what by when, what breaks first, what's the cheapest experiment to learn.
• The Empath: stakeholder feelings, narrative, who gets hurt, how this lands publicly.
• The Compliance Voice: regulatory exposure, audit trail, legal risk, "what does counsel say."
• The Customer Whisperer: what real users actually do (not what they say in surveys).
• The Contrarian: argues the opposite of consensus by default, on principle.
• The Engineer: technical risk, build complexity, what's actually shippable.
• The Capital Allocator: opportunity cost, where else this dollar or this hour goes.
• The Veteran: pattern-matches to a prior similar decision and tells you what happened.

For domain-heavy questions (fintech, telecom, healthcare, M&A) consider swapping in a domain specialist (e.g. a Card Network Veteran for an interchange decision, a Spectrum Lawyer for a telecom build). Friction beats labels. If two advisors would say the same thing, replace one.

Round 1 prompt template

For each advisor, spawn a subagent with this exact prompt structure:

You are <ADVISOR NAME>. <ONE-LINE LENS DESCRIPTION>.

The decision: <RESTATED QUESTION>
What we know: <USER-PROVIDED CONTEXT, verbatim, no editorializing>

Write a 150-250 word position from your lens. End with a single sentence starting "Recommendation:" that names a direction.

Rules:
- Stay in your lens. Do not play the other advisors.
- Do not summarize the question back at the user.
- Do not hedge. If you are uncertain, say what would resolve it.
- No preamble. Get to the point in the first sentence.
- No em dashes anywhere. Use commas, parens, colons, or sentence breaks.
<EVIDENCE_REQUIREMENT, only present for evidence-claiming advisors>

Evidence-claiming advisors must cite

Some advisor lenses derive authority from citation, not reasoning. Examples: Doc Reader, Compliance Voice, Card Network Veteran, Spectrum Lawyer, Auditor, anyone whose value-add is "what does the source say." These advisors MUST do tool use (WebFetch, Read, Grep) and quote specific URLs, file paths, or line numbers in their position. A Doc Reader who recites from training is just a Strategist with worse epistemics; the persona collapses without live evidence.

When spawning an evidence-claiming advisor, append this to their Round 1 prompt:

EVIDENCE REQUIREMENT (mandatory):
You must do tool use before writing your position. Use WebFetch, Read, or Grep to retrieve the source you are citing. Quote at least two specific facts with their exact URL or file path. If a fetch fails or contradicts your prior belief, say so explicitly. A position with zero tool uses fails this lens; rewrite as a different advisor type or report "I cannot make this case without source access."

The Chairman must verify tool use before counting an evidence-claiming advisor's position. If a Doc-Reader-style advisor produced 0 tool uses, downgrade their peer-review scores by 1 point on Sharpness AND Risk awareness (you cannot be sharp about facts you did not check, and you cannot see the risk of being wrong about an unverified claim). If their position becomes load-bearing for the verdict, respawn them with the evidence requirement before publishing.

Round 2 prompt template

Take the 5 positions. Strip names. Label them Advisor A, B, C, D, E. For each advisor, spawn a subagent with the four positions they did not write:

You are <ADVISOR NAME>. Below are four anonymous positions from your peers on the same question. You do not know who wrote which.

Grade each on a 1-5 scale across:
- Sharpness: is the recommendation crisp, or is it mush?
- Relevance: does it actually help decide the question, or is it interesting-but-tangential?
- Risk awareness: does it see what could go wrong?
- Actionability: can the user do something with this Monday morning?

For each peer (A through D in your view), output:
- Scores: Sharpness X, Relevance X, Risk X, Actionability X
- Critique: one sentence, honest. Politeness is a tax on the user.

Aggregate the scores into a scoreboard: mean per advisor across the 4 dimensions, plus a rank. The user wants to see who carried the day and who got dragged.

Chairman's verdict: output structure

The Chairman's job is to deliver a strong recommendation, period. The friction the council produced is the input. The verdict is a single direction the user can act on, with the strongest dissent called out so they know what to watch for. Hedging is failure. A weak verdict is worse than a wrong one because it shifts the decision back to the user, which is exactly what this skill exists to prevent.

Use this exact structure. Verdict first, council debate collapsed below.

## Verdict
<One paragraph. Pick a direction in the FIRST sentence. Name the strongest dissent in one clause. No fence-sitting.>

## Why this, not the alternative
<2-4 bullets of load-bearing reasoning. Cite advisors by role, not letter, e.g. "the Quant flagged...". Do not introduce new arguments the council did not make.>

## What would change my mind
<2-3 specific signals or facts that would flip the verdict. Concrete and observable, not "if circumstances change.">

## Next step
<ONE action. Today or this week. Specific enough that the user knows exactly what to open, write, run, or who to message. "Draft the term sheet" is good. "Think about it more" is not.>

## Council scoreboard
| Advisor | Sharpness | Relevance | Risk | Actionability | Mean |
|---|---|---|---|---|---|
| ... | ... | ... | ... | ... | ... |

<details>
<summary>Round 1: full positions</summary>

### <Advisor 1>
<position text>

### <Advisor 2>
...
</details>

<details>
<summary>Round 2: peer critiques</summary>
<critiques organized by reviewer or by reviewee, your choice. Keep tight.>
</details>

How to read the council vote

The Chairman aggregates positions, not just scores. After Round 2, classify the split:

• 5-0 or 4-1: clear majority. The Chairman picks the majority direction. The minority view is minority-reported (named in the verdict in one clause, expanded in "What would change my mind" as a future signal). It is NEVER promoted into the plan or hedged into the verdict.
• 3-2: real split. Pick the side with the higher peer-review mean on the Sharpness + Actionability axes (those carry decisions; Risk and Relevance are softer). State the split plainly in the verdict, then pick.
• 2-2-1: three-way split. Pick the position that two converging advisors share, name the other two camps as dissent, and tell the user what extra information would tip it. Still pick.

The Chairman never reports a split without picking. "The council was divided" is not a verdict. The user can read the scoreboard themselves.

Banned patterns in the verdict

The verdict paragraph fails if it contains any of these:

• The words "if", "then", "first", "before", "pending", "depends on", or "assuming". These smuggle conditional plans into a verdict that should be a direction.
• Multi-step language: "do A, then B, then C". A verdict is one decision. The Next step section names ONE action that follows from it.
• "Gather data first" or "run an experiment before committing". If a measurement is genuinely cheap, it goes in the Next step. It is not the verdict.
• "On balance" or "ultimately" or "it depends" or "this is your call". Delete and rewrite.
• Synthesizing the minority into the plan ("execute the audit, but with the Skeptic's caveat that..."). The minority gets one clause for context and a slot in "What would change my mind". Not a slot in the verdict.

Banned patterns in Next step

• Numbered lists, bulleted lists, or "and also". If you have two actions, you have not picked the one that matters.
• Conditional branches ("if X then do Y, else Z"). The point of a Next step is that the user does not have to decide what comes first.
• Vague verbs: "explore", "investigate", "consider", "think about", "review". Use concrete verbs: open, write, run, message, draft, ship, delete, merge, file.

When the council genuinely cannot pick

Rare. Reserved for cases where two camps share equal peer-review means AND each is conditioned on different external facts the user has not provided. In that case ONLY: state the two directions plainly in the verdict, name the single fact that would resolve them, and put "answer that fact" as the Next step. Do not use this as an escape hatch when one camp is actually winning.

Style rules

• Each advisor is a person with voice and edge, not a panel of experts. Give them an opening line that sounds like them.
• The Skeptic and the Quant will often clash with the Empath. Do not smooth this over. The friction is the product.
• The Chairman picks. Always. The verdict's first sentence is a direction the user can act on. "Yes, do X" or "No, do Y instead." Not "do part of X, gather data, then decide."
• The Chairman names the strongest dissent in one clause inside the verdict, then expands it as a future signal in "What would change my mind". The dissent is NEVER promoted into the plan or the Next step. Synthesizing the minority view into the verdict is the failure mode this skill exists to prevent.
• The Chairman never says "ultimately, this is your call," "on balance," "it depends," or "this is a judgment call." Delete and rewrite as a direction.
• 150-250 words per position is a ceiling, not a floor. If an advisor lands their case in 80 words, ship it.
• The scoreboard is numbers and a table. No paragraph commentary inside it.
• No em dashes anywhere in the output. Use commas, parens, colons, or full stops. This applies to advisors, peer reviews, and the Chairman.
• No corporate fluff: "leverage," "synergies," "robust," "holistic." Speak like a real person who has skin in the game.

Self-check before shipping the verdict

Before posting the response, the Chairman re-reads the verdict paragraph and the Next step against this checklist. If any answer is "yes," rewrite before shipping.

1. Does the verdict contain "if", "then", "first", "before", "pending", "depends on", "assuming", or "on balance"?
2. Is the verdict more than one direction? (Two directions stitched together with "and" still counts as hedging.)
3. Did I synthesize the minority view into the plan or the Next step?
4. Is the Next step a list, a checklist, or a multi-step sequence?
5. Could a reader finish the verdict and still not know what I am telling them to do?
6. Did I cast an evidence-claiming advisor (Doc Reader, Compliance Voice, domain specialist, Auditor) who produced 0 tool uses, AND is their position load-bearing for the verdict? If yes, the verdict's authority is borrowed; respawn that advisor with the evidence requirement, or remove their reasoning from the verdict.

Six "no" answers means ship. Even one "yes" means rewrite.

Why this works

A single LLM answer to a decision question tends toward the bland centroid: balanced, cautious, mealy. Five voices forced into distinct lenses produce real disagreement. Anonymous peer review punishes empty rhetoric (it cannot hide behind a famous name) and rewards crisp, actionable thinking. The Chairman's job is to actually decide, which is the part most LLM workflows skip. The user gets a verdict they can act on plus the dissent they need to watch for.