Consensus Voting
hjjtt/claude-skill-consensus-voting
Múltiplos modelos, lentes diversas, síntese por consenso
Ver no GitHubSkill que implementa votação adversarial multi-modelo para decisões críticas. Executa N votantes independentes com lentes analíticas distintas em paralelo, depois sintetiza via juiz para identificar consenso (sinal) e desacordo (incerteza quantificável).
Instalação
npx convez-ai hjjtt/claude-skill-consensus-voting1 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
- Previsões em esportes e competições
- Análise de investimentos e decisões financeiras
- Arquitetura de código e decisões técnicas
- Análise de contratos e revisão legal
- Diagnóstico e raciocínio crítico em domínios especializados
Veredito

Abordagem sofisticada para reduzir viés de modelo único em decisões irreversíveis. Ideal para quem precisa de confiança além da resposta pontual — exige orçamento de tokens e latência para múltiplas chamadas paralelas.
README
A tradução para o português deste README ainda está na fila.
Consensus Voting — Claude Code Skill
Multi-model adversarial agreement for high-stakes decisions. Spin up N independent voters with different analytical lenses, then synthesize via a judge.
✨ Why this skill exists
A single LLM has blind spots. Ask one model to predict a match, audit a contract, or pick an architecture, and you get that model's bias — not the truth.
Consensus Voting runs multiple models with deliberately different analytical lenses in parallel, then has a judge synthesize where they agree (signal) and where they disagree (uncertainty to quantify).
Agreement across diverse models = signal. Disagreement = uncertainty you should not paper over.
🎯 When to use
| Scenario | Why consensus helps |
|---|---|
| Sports / match prediction | Form, injuries, altitude — no single lens weights them all correctly |
| Investment & financial analysis | Bull vs bear lenses catch different risks |
| Code architecture decisions | Performance / maintainability / DX prevent blind spots |
| Medical / diagnostic reasoning | Cross-check reduces anchoring bias |
| Legal / contract review | Different lenses catch different clause risks |
| Strategic / competitive analysis | Opponent modeling benefits from diverse assumptions |
| Any irreversible decision | When being wrong costs more than running 3 extra models |
Trigger keywords: high accuracy, multi-model, consensus, voting, compare models, predict, forecast, 2串1, parlay, betting, verify accuracy.
🏗️ Architecture
┌─────────────────────┐
│ User Question │
└─────────┬───────────┘
│
┌──────────────┼──────────────┐
│ │ │
┌─────▼─────┐ ┌─────▼─────┐ ┌─────▼─────┐
│ Voter A │ │ Voter B │ │ Voter C │
│ (Opus) │ │ (Sonnet) │ │ (Haiku) │
│ Deep logic │ │ Data-driven│ │ Fast gut │
└─────┬─────┘ └─────┬─────┘ └─────┬─────┘
│ │ │
└──────────────┼──────────────┘
│
┌─────────▼───────────┐
│ Judge (Opus) │
│ Synthesize + Rank │
└─────────┬───────────┘
│
┌─────────▼───────────┐
│ Final Consensus │
│ + Confidence Score │
└─────────────────────┘
The skill enforces lens diversity: each voter argues from a distinct analytical axis (data / momentum / signal / narrative / psychology). Identical prompts produce correlated outputs, which defeats the entire purpose.
🚀 Installation
Drop the skill into your Claude Code skills directory:
macOS / Linux
git clone https://github.com/hjjtt/claude-skill-consensus-voting.git \
~/.claude/skills/consensus-voting
Windows (PowerShell)
git clone https://github.com/hjjtt/claude-skill-consensus-voting.git `
$env:USERPROFILE\.claude\skills\consensus-voting
Restart Claude Code. The skill is now available to all sessions.
💡 Usage
Just ask Claude in natural language — the skill auto-activates on trigger keywords:
> Predict the World Cup quarterfinals with high accuracy
> Use consensus voting to evaluate these 3 architecture options
> Run a multi-model check on this contract clause
> 2串1 最稳方案
Or invoke explicitly:
> /consensus-voting Should I migrate to Postgres or stay on MySQL?
Example — Sports Prediction (Real Output)
📊 MULTI-MODEL CONSENSUS RESULT
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🗳️ Vote Tally:
Switzerland Win: ████ 4/4 models
Draw: ░░░░ 0/4 models
Qatar Win: ░░░░ 0/4 models
🎯 Consensus: Switzerland 2-0 @ 8.4/10 confidence
Probability: SUI 77% / Draw 16% / QAT 7%
⚠️ Key Risk: Switzerland rotation if group already secured
🔑 Pivot Variable: Whether Xhaka starts
🧪 The 4-Step Pattern
Step 1 — Context Harvest
Parallel research agents gather raw data. Each covers a different dimension.
Step 2 — N-Model Independent Vote
3-5 voters in parallel. Each gets the same data but a different analytical lens:
opus— Deep structural analysis, multi-factor reasoningsonnet— Balanced data-driven analysishaiku— Fast signal extraction, top-3 factors onlyfable— Narrative, creative, psychological angles
Step 3 — Judge Synthesis
A single judge reads all votes and produces:
- CONSENSUS ITEMS — what ≥75% of voters agree on
- DISSENT ITEMS — where voters diverge, and why
- FINAL RECOMMENDATION — with explicit 1-10 confidence
- RISK MATRIX — what could invalidate the consensus
- KEY VARIABLE — the single info that, if known, would most change the answer
Step 4 — Present Results
Vote tally + consensus + confidence + risks, in scannable format.
📐 Lens Design Rules
Quality depends on voter diversity, not voter count:
- No two voters share the same analytical axis — redundant voters add cost without signal
- Each lens must be defensible in isolation — no "just guess" voters
- Match model tier to lens depth — don't run 5 Opus voters when 1 Haiku covers fast pattern matching
- Minimum 3 voters — below 3, you can't distinguish agreement from coincidence
- Odd numbers prevent ties — use 3 or 5; if 4, judge breaks ties
📊 Confidence Calibration
| Confidence | Meaning | Action |
|---|---|---|
| 9-10 | Near-unanimous + strong structural factors | Proceed without hedging |
| 7-8 | Majority consensus, moderate dissent | Proceed with stated risks |
| 5-6 | Split vote, single variable dominates | Hedge or wait for pivot info |
| 3-4 | No clear consensus | Present options, don't recommend |
| 1-2 | Models actively contradict | Flag as unpredictable, refuse to recommend |
💰 Cost Awareness
| Setup | Agents | Approx. Cost |
|---|---|---|
| Minimal (3 voters + 1 judge) | 4 + research | ~$0.15–0.30 |
| Standard (4 voters + 1 judge) | 5 + research | ~$0.25–0.50 |
| Exhaustive (5 voters + 2 research + 1 judge) | 8 | ~$0.50–1.00 |
Don't cheap out on high-stakes calls. Don't run 10 agents on a $5 question.
⚠️ Anti-Patterns
| ❌ Bad | ✅ Good |
|---|---|
| Same prompt to all models | Different analytical lenses per voter |
| 5 voters, all Opus | Mix model tiers to match lens complexity |
| Judge ignores minority dissent | Judge explains why dissent was overruled |
| No confidence score | Always output calibrated confidence |
| One-shot, no research phase | Always harvest context before voting |
| Present consensus as fact | Always present with risk matrix |
🛠️ Requirements
- Claude Code ≥ recent version with Workflow tool
- Access to multiple Claude model tiers (Opus / Sonnet / Haiku / Fable)
📄 License
MIT — use freely, attribute kindly.
🙏 Credits
Authored with Claude Code. Architecture inspired by adversarial agreement and ensemble methods.
If this skill saved you from a bad single-model call, consider starring the repo ⭐