convez.ai
SkillProdutividadeIntermediário
unhingged

Council — Conselho de Decisão com Subagentes

unhingged/council

Pressão-testa decisões com 5 advisors agentic independentes + Chairman

Ver no GitHub
BomPluginScore · de 0 a 100

Uma skill que submete decisões a um conselho agentic de cinco advisors independentes (Contrarian, First Principles, Expansionist, Outsider, Executor), cada um rodando como subagente isolado, seguido de um veredito final do Chairman. Pressão-testa decisões através de perspectivas conflitantes e paralelas, evitando o viés de consenso artificial.

Instalação

$npx convez-ai unhingged/council

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
87%
Último commit
há 6 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Validar decisões estratégicas antes de implementação
  • Explorar múltiplas perspectivas em problemas complexos
  • Reduzir viés de confirmação em análises críticas
  • Debater argumentos técnicos ou de negócio entre posições opostas

Veredito

Council é uma abordagem sofisticada para contrariar o viés de consenso artificial em análises. A arquitetura de subagentes isolados resolve um problema real — single-session role-play tende a ser cosmética — e o framework de cinco advogados (contrarian, expansionist, executor etc.) fornece estrutura prática. Requer Claude Code com suporte a subagentes; sem dependências externas.

README

A tradução para o português deste README ainda está na fila.

A Claude Code skill that runs questions through an agentic council of five advisors. Each advisor runs as a separate subagent with its own isolated context. They argue in parallel, optionally over multiple debate rounds, and a Chairman subagent renders the verdict.

Why subagents

A single Claude session role-playing five voices knows what the other voices will say while it's writing each one, so the "disagreement" tends to be cosmetic — five lanes that quietly complement each other.

This skill spawns the advisors as independent subagents with no shared context. None of them sees the others until the (optional) debate rounds, and even then only an anonymized, shuffled version. They can contradict each other, argue from incompatible premises, and refuse to converge — which is the whole point of pressure-testing.

In blind A/B/C testing (an independent model judged anonymized outputs without knowing which was which), the agentic approaches — subagents and a separate claude -p orchestrator — both beat a single-session simulation on verdict quality and usefulness, and tied each other. Subagents were lighter and avoided a context-bleed the subprocess approach occasionally showed, so subagents are the engine.

The advisors

AdvisorRole
The ContrarianFinds the fatal flaw
The First Principles ThinkerAsks if you're solving the right problem
The ExpansionistFinds the upside everyone missed
The OutsiderCatches the curse of knowledge
The ExecutorCares only what you do Monday morning

The Chairman runs last, sees the full transcript, and renders a structured verdict. The Chairman is a pure judge — they weigh and pick, they don't add new arguments.

Architecture

flowchart TD
    Q(["User question"]) --> F["Frame neutrally<br/>framed.txt"]

    subgraph advisors ["5 advisor subagents · parallel · isolated context"]
        direction LR
        A1["The Contrarian"]
        A2["First Principles"]
        A3["The Expansionist"]
        A4["The Outsider"]
        A5["The Executor"]
    end

    F --> A1 & A2 & A3 & A4 & A5
    A1 & A2 & A3 & A4 & A5 --> C["Chairman subagent<br/>verdict.md"]
    C --> S[["session.md"]]

    classDef frame fill:#5b2a86,stroke:#f4f1e8,color:#ffffff;
    classDef advisor fill:#e0623e,stroke:#f4f1e8,color:#ffffff;
    classDef judge fill:#1b1b1f,stroke:#f4f1e8,color:#f4f1e8;
    classDef out fill:#f4f1e8,stroke:#5b2a86,color:#1b1b1f;
    class F frame
    class A1,A2,A3,A4,A5 advisor
    class C judge
    class S out

Optional debate rounds (2–3): each advisor runs again seeing the others' previous round — anonymized and shuffled — so they engage with arguments, not authors. Then the orchestrator stitches the files into session.md.

Requirements

  • Claude Code with subagent support

No scripts, no Python, no dependencies — the skill is instructions plus six advisor/Chairman personas.

Installation

This repo is both a Claude Code plugin marketplace and an Agent Skill, so it installs two ways.

Claude Code plugin (/plugin):

/plugin marketplace add unhingged/council
/plugin install council@unhingged

Agent Skills / skills.sh (npx skills, works across Claude Code, Cursor, and other agents):

npx skills add unhingged/council --skill council

Both resolve to the same single source at skills/council/SKILL.md — there's no duplicated copy to keep in sync. (--skill council is needed because the skill lives in the skills/ subdirectory rather than the repo root, which is what lets the same tree also be a Claude Code plugin.)

Manual / project-local: copy skills/council/ into your ~/.claude/skills/ (personal) or .claude/skills/ (project) directory.

Repository layout

council/
├── .claude-plugin/
│   ├── plugin.json          # declares the "council" plugin
│   └── marketplace.json     # marketplace "unhingged" listing the plugin (source ".")
├── skills/council/
│   ├── SKILL.md             # the skill — read by both install channels
│   └── prompts/             # the six advisor + Chairman personas
├── README.md  LICENSE

Usage

Invoke explicitly. The skill is heavy by design and only triggers when you ask:

> Council this: should I quit my job to start a company?

The skill (see SKILL.md) walks the orchestrating Claude through: framing the question, dispatching the five advisor subagents in parallel, optionally running debate rounds (with the previous round anonymized and shuffled so advisors react to arguments, not authors), dispatching the Chairman, and stitching the files into session.md.

Output structure

Every session produces session.md with:

  1. ## The question — neutral restatement
  2. ## The advisors (per round) — five independent responses
  3. ## Verdict — Chairman ruling: what survived / what didn't / most right / recommendation / a note on the council

Credit

The original "LLM Council Facilitator" framing and the five advisor roles come from a publicly circulated prompt concept. This skill adds the agentic subagent orchestration, configurable debate rounds with anonymized cross-feed, the pure-judge Chairman pattern, and the output contract.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados