LLM Council — Skill de Decisão com 5 Conselheiros IA
Tyrusrechs11/llm-council-skill
5 conselheiros IA debatem sua decisão, pontuam-se e sintetizam um veredito confiável
Ver no GitHubSkill para Claude Code que submete decisões, planos ou código a 5 conselheiros IA com perspectivas distintas, que se avaliam anonimamente e geram um veredito com confiança explícita. Inclui detecção de convergência, peer review pontuado e relatório HTML estruturado.
Instalação
npx convez-ai Tyrusrechs11/llm-council-skill1 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
Casos de uso
- Validar decisões estratégicas (pivôs de negócio, mudanças de posicionamento)
- Pressure-test planos e arquiteturas de sistema antes de implementar
- Gerar múltiplas perspectivas em problemas complexos de startup/produto
Veredito

Skill sólido e bem documentado que traz rigor real ao processo de tomada de decisão multi-perspectiva. Ideal para quem precisa validar ideias complexas antes de agir, mas exige contexto rico para não virar apenas ruído estruturado.
README
A tradução para o português deste README ainda está na fila.
LLM Council — Claude Code Skill
Stop trusting one perspective. Run any decision, plan, or code through 5 AI advisors who argue from different angles, score each other anonymously, and hand you a verdict you can actually trust.
Enhanced fork of @tenfoldmarc's llm-council-skill, based on Andrej Karpathy's LLM Council methodology.
What it does
Type council this: [your question] in Claude Code and you get:
- 5 advisors analyse in parallel — each locked into a distinct thinking style (Contrarian, First Principles, Expansionist, Outsider, Executor)
- Each advisor ends with a Signal block — a one-line verdict, confidence rating (1–10), and "what would shift my view"
- 5 peer reviewers score them anonymously — Rigour and Novelty (1–5), strongest response, biggest blind spot, what all five missed
- Convergence check — are they aligned, split, or polarised? If strongly aligned, the chairman steelmans the opposite
- Chairman synthesises — where they agree, where they clash, what peer review surfaced, a clear recommendation with confidence rating, one concrete next step
- HTML report + markdown transcript saved to
~/council-sessions/
Total time: ~3–5 minutes.
What this adds over the original
| Feature | Original | This fork |
|---|---|---|
| Advisor output | Free-form text | Structured Signal block (verdict + confidence + "what would shift my view") |
| Peer review | 3 qualitative questions | Scored per response (Rigour/Novelty 1–5) + strongest/gap/missed |
| Convergence detection | None | Flags consensus, split, or polarised — steelmans opposite when strongly converged |
| Chairman confidence | None | Explicit X/10 rating + "if I'm wrong it's because..." |
| Domain detection | None | Appends tech/creative lens to advisor prompts when relevant |
| HTML report | Vague "clean styling" | Convergence bar, colour-coded confidence badges, scores table, native collapsibles |
| Output location | Workspace root | Dedicated council-sessions/ folder |
Install
git clone https://github.com/Tyrusrechs11/llm-council-skill ~/.claude/skills/llm-council
Restart Claude Code. The skill loads automatically.
Use
Trigger phrases:
council this: [question]run the council on: [question]pressure-test this: [idea]stress-test this: [plan]war room this: [decision]debate this: [question]
Example:
council this: I'm thinking of pivoting from a £297 course to a £97 live workshop. Is that the right move?
Give it context — the richer the input, the sharper the output.
Good council questions
- "Should I launch X or Y?"
- "Which of these 3 positioning angles is strongest?"
- "I'm thinking of pivoting from X to Y — am I crazy?"
- "Here's my system architecture — what's the weakest point?"
- "Should I hire a VA or build an automation first?"
Skip the council for: factual questions, pure creation tasks ("write me a tweet"), or when you've already decided and just want validation.
Output
Every session produces two files in ~/council-sessions/ (or ./council-sessions/ if in a project):
council-report-[YYYY-MM-DD-HHmm].html # visual report, opens automatically
council-transcript-[YYYY-MM-DD-HHmm].md # full transcript for reference
Honest framing
This is a structured multi-perspective reflection framework — one model adopting constrained personas, applying explicit scoring criteria, and synthesising across them. The advisors share a context window; they are not truly independent processes. The value comes from the structured constraints forcing distinct angles, not from epistemic independence.
Credit
- Methodology: Andrej Karpathy's LLM Council
- Original Claude Code adaptation: @olelehmann
- Original skill: @tenfoldmarc
- This enhanced fork: @Tyrusrechs11
Licence
MIT — see LICENSE.