convez.ai
SkillDados & AnalyticsAvançado
hjjtt

Consensus Voting

hjjtt/claude-skill-consensus-voting

Múltiplos modelos, lentes diversas, síntese por consenso

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill 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-voting

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
5
Forks
0
Acompanhando
1
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
2

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.

Claude Code License: MIT


✨ 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

ScenarioWhy consensus helps
Sports / match predictionForm, injuries, altitude — no single lens weights them all correctly
Investment & financial analysisBull vs bear lenses catch different risks
Code architecture decisionsPerformance / maintainability / DX prevent blind spots
Medical / diagnostic reasoningCross-check reduces anchoring bias
Legal / contract reviewDifferent lenses catch different clause risks
Strategic / competitive analysisOpponent modeling benefits from diverse assumptions
Any irreversible decisionWhen 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 reasoning
  • sonnet — Balanced data-driven analysis
  • haiku — Fast signal extraction, top-3 factors only
  • fable — Narrative, creative, psychological angles

Step 3 — Judge Synthesis

A single judge reads all votes and produces:

  1. CONSENSUS ITEMS — what ≥75% of voters agree on
  2. DISSENT ITEMS — where voters diverge, and why
  3. FINAL RECOMMENDATION — with explicit 1-10 confidence
  4. RISK MATRIX — what could invalidate the consensus
  5. 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:

  1. No two voters share the same analytical axis — redundant voters add cost without signal
  2. Each lens must be defensible in isolation — no "just guess" voters
  3. Match model tier to lens depth — don't run 5 Opus voters when 1 Haiku covers fast pattern matching
  4. Minimum 3 voters — below 3, you can't distinguish agreement from coincidence
  5. Odd numbers prevent ties — use 3 or 5; if 4, judge breaks ties

📊 Confidence Calibration

ConfidenceMeaningAction
9-10Near-unanimous + strong structural factorsProceed without hedging
7-8Majority consensus, moderate dissentProceed with stated risks
5-6Split vote, single variable dominatesHedge or wait for pivot info
3-4No clear consensusPresent options, don't recommend
1-2Models actively contradictFlag as unpredictable, refuse to recommend

💰 Cost Awareness

SetupAgentsApprox. 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 modelsDifferent analytical lenses per voter
5 voters, all OpusMix model tiers to match lens complexity
Judge ignores minority dissentJudge explains why dissent was overruled
No confidence scoreAlways output calibrated confidence
One-shot, no research phaseAlways harvest context before voting
Present consensus as factAlways 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 ⭐

Ver o original no GitHub

Bom para

Relacionados