convez.ai
SkillProdutividadeIntermediário
nemock

Council Review

nemock/council-review

Múltiplas perspectivas paralelas + síntese estruturada de achados

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que reúne um conselho de revisores especialistas independentes para pressure-test paralelo de projetos, sistemas ou documentos, sintetizando um relatório único priorizado. Cada revisor trabalha cegamente sem influência dos outros, surfando consensos e tensões interessantes.

Instalação

$npx convez-ai nemock/council-review

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Pressure-test de arquitetura ou PRDs antes de validação
  • Red-team de estratégias (pricing, produto, mercado)
  • Review multi-ângulo de documentos, rotinas ou mudanças antes de merge

Veredito

Ferramenta sofisticada para quem precisa de análise real multi-perspectiva sem viés de ancoragem. Ideal para decisões não-triviais onde visão única é risco; exige planejamento de tokens e latência (60–180s).

README

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

council-review

A Claude Code skill that convenes a council of independent expert reviewers to pressure-test any project, system, document, or plan — in parallel — then synthesizes one prioritized report.

License: MIT

A single reviewer produces a coherent but narrow read of any non-trivial artifact. Running reviewers sequentially does not fix this — each one anchors on the previous one's framing. council-review runs 3–6 lenses in parallel and blind, then a chair synthesizes the findings, surfacing where lenses agreed (high-confidence) and where they conflicted (interesting tensions).

It is report-only. It never modifies the subject under review.


Install

The skill is a single SKILL.md file. Drop the repo into your skills directory:

git clone https://github.com/nemock/council-review.git ~/.claude/skills/council-review

Then restart Claude Code (or start a new session). The skill will appear under its name in the available-skills list.

To update later: git -C ~/.claude/skills/council-review pull.

To uninstall: rm -rf ~/.claude/skills/council-review.


Use

Trigger the skill by asking for a multi-perspective review:

/council-review the routine in scheduled_routines/founder_tip_tuesday/
Run a council review on my landing page draft at /tmp/landing.md — be thorough.
Pressure-test the architecture in PRD.md from 4 angles: market, tech, ops, risk.

You can also let Claude pick the skill from natural-language phrasing:

Get me multiple perspectives on this PR before I merge it.
Red-team my pricing strategy.
Review this from different angles.

The skill will scope the subject, select lenses, fan out reviewers, and return a single synthesis report.


What you get

# Council Review — <your subject>

## Verdict
<overall health, 2-4 sentences>

## Council
<one line per lens with its headline take>

## Strengths to preserve
- ...

## Cross-cutting findings (multiple lenses agreed)
- ...

## Conflicts & tensions
- <where lenses disagreed and the chair's call>

## Prioritized recommendations
1. **<change>** (severity, lens) — what to change and why it matters.

## Per-lens detail
<per-lens summaries preserved for the depth reader>

The verdict + prioritized recommendations are designed to be skimmed and acted on. The per-lens detail is there when you want the raw read.


How it works

The skill drives Claude Code's Workflow tool to fan out one subagent per lens — each in its own context, each blind to the others — and returns the findings as structured JSON. The invoking session then acts as the chair: it reads all findings and writes one report following fixed synthesis rules (agreement is signal, conflicts are not averaged, severity is re-ranked holistically, strengths are named).

The skill ships with four canonical lens packs (content/marketing routine, software/codebase, product/strategy, document/writing) and adapts them to the subject. You can also name your own lenses.

See SKILL.md for the full mechanism and PRD.md for design rationale.


Cost note

Each lens is a full subagent context. Five lenses on a moderate artifact is a non-trivial token spend (and a 60–180s wall clock). Scale lens count to the decision's importance:

  • 3 lenses — quick pressure test.
  • 4 lenses — standard review.
  • 5–6 lenses — "thorough" or "red-team."

The skill won't go above 6 lenses without an explicit ask.


What this is NOT

  • Not a code-fixer. The skill recommends; it never edits. If you want changes applied, that's a separate explicit step.
  • Not a single-lens review. For a sharp single-perspective read, use the corresponding focused skill (e.g. /code-review, /security-review).
  • Not deep research. Reviewers evaluate the provided subject only — they don't go fetch external sources. For research, use /deep-research.

Attribution

Inspired by Andrej Karpathy's llm-council. His project runs in a very different way — it spins up a web app that queries several different vendor models (OpenAI, Anthropic, Google, xAI) and has them rank each other's answers. That was a bit beyond what I was looking for. I was okay with the different personas all coming from Claude. What I wanted was something simpler — an internal red team that gives me distinct angles on the same artifact without me having to set up multi-vendor billing or run a server.

I do understand the tradeoff. Different lenses inside the same model family will produce more correlated answers than a true multi-vendor council, so some of the divergence Karpathy's setup buys you isn't available here. In practice, the structured lens prompts plus the explicit "find conflicts, don't average them" synthesis rule recover most of the value — at least for the kinds of reviews I run.


License

MIT — use it, fork it, ship it.

Ver o original no GitHub

Bom para

Relacionados