LLM Council (5-persona debate)

SKILL.md

---
description: Run a 5-persona LLM Council debate (independent first opinions, anonymized peer review and ranking, then a chairman synthesis) to choose the strongest option before producing the final output. Adapted from Andrej Karpathy's LLM Council. Trigger phrases: run a council, llm council, debate this with 5 personas, council of personas to pick the best angle, get multiple perspectives before deciding.
---

# LLM Council (5-persona debate)

Pick the strongest option, or produce the most defensible answer, by running a council of 5 conflicting personas, having them critique and rank each other, then synthesizing with a chairman. Adapted from Andrej Karpathy's LLM Council (github.com/karpathy/llm-council): the same three stages (independent first opinions, anonymized peer review, chairman synthesis), run on one model with 5 role-played personas instead of 4 separate models, so it works in a single agent session with no extra API keys.

## When to use

A hard, subjective, or high-stakes "which of these is best" or "what is the right take" call, where one pass would be shallow or biased. Choosing a content angle (which reel hook to run), a strategy, a name, a priority. Skip it for simple factual or mechanical tasks.

## Inputs

- The question, or the set of options to evaluate.
- Optional: the 5 personas (else use the defaults below).
- Optional: the chairman (default: a neutral synthesizer).

## The council (5 conflicting personas)

Use 5 genuinely opposing viewpoints so the debate has friction. General-purpose defaults:

1. Maximalist, pushes the boldest, highest-upside version.
2. Skeptic, attacks the weakest assumption and finds where it fails.
3. Operator, only cares what is practical and shippable now.
4. Audience, speaks for the end user or viewer and what they actually feel.
5. Contrarian, argues against the obvious consensus.

For an IG reel angle, swap to: hook-maximalist, the algorithm (what the feed rewards), the niche viewer (the specific person you are talking to), the brand voice, the skeptic.

## Procedure

Stage 1, first opinions (independent). For each persona, in that persona's voice, give an independent answer, or an independent pick plus reason among the options. The personas do not see each other yet. Keep each to a few sentences.

Stage 2, anonymized review. Relabel the five opinions A to E with the persona names stripped, so no one can play favorites. Then each persona reads all five and ranks them strongest to weakest on accuracy, insight, and likely real-world impact, with one line of reasoning per rank. Aggregate into a consensus order.

Stage 3, chairman synthesis. The chairman reads the five opinions and the consensus ranking, then writes the final answer: the chosen option, the two or three reasons it won, and the single strongest objection that survived. Merge the best of the top two if that beats either alone. Do not just rubber-stamp the top rank.

## Decision points

- If the first opinions already agree, say so and shortcut straight to the chairman.
- If the ranking splits with no clear winner, the chairman names the split and makes the call with a stated reason, no hedging.
- If two personas converge, replace one with a sharper opposing view. The value is in the friction.

## Output contract

1. The final pick or answer, one short paragraph.
2. Why it won, two or three bullets.
3. The best surviving objection, one line, so the tradeoff stays visible.
4. Optional appendix: the five anonymized opinions and the consensus ranking, for inspection.

If the output is human-facing copy, hand the chairman's text to a humanizer pass before shipping. The council debates the idea, it does not polish the prose.

## Outcome signal

Whether the chosen option actually performed (the reel angle that got the most reach, the strategy that worked). Feed that back so the council's picks can be judged over time.

## Credit

Pattern by Andrej Karpathy, "LLM Council" (github.com/karpathy/llm-council). This is a single-model, persona-based adaptation.