convez.ai
AgenteDados & AnalyticsAvançado
zpower426

DataPowers

zpower426/datapowers

Protocolo de auditoria estatística para pipelines de dados robusto e à prova de vazamentos

Ver no GitHub
InicialPluginScore · de 0 a 100

Protocolo profissional de mineração de dados e ML para agentes de IA que implementa rigor estatístico, detecção de vazamento de dados e validação em três camadas. Garante integridade dos dados através de auditoria sistemática antes da análise, sem depender apenas de métricas de modelo.

Instalação

$npx convez-ai zpower426/datapowers

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
57%
Último commit
há 7 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Validação rigorosa de dados antes de treinamento de modelos
  • Detecção e prevenção de data leakage em pipelines ML
  • Análise estatística com intervalos de confiança e testes de significância
  • Workflows hypothesis-driven para ciência de dados

Veredito

DataPowers é um framework sólido e bem estruturado para quem trabalha com ML em produção e precisa de garantias reais sobre integridade dos dados. A abordagem baseada em protocolos de auditoria (não apenas prompts) é diferenciada, mas exige disciplina e conhecimento técnico profundo para implementar.

README

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


datapowers is a professional data mining and machine learning workflow protocol for Claude Code, Gemini CLI, and autonomous agents. It bridges software engineering rigor (TDD) with data science's statistical sovereignty — enforcing leakage prevention, three-layer validation, and evidence-based delivery as non-negotiable pipeline gates.

🔍 Why datapowers?

Statistical Integrity > Model Metrics. datapowers is not a prompt collection — it is an Audit Protocol.

  • 🛡️ Leakage Defense — Systematic temporal and preprocessing leakage audits at the source, before any evaluation.
  • ⚖️ Three-Layer Validation — Physical (Schema), Logical (Business Rules), and Statistical (Distribution Drift) gates block training on bad data.
  • 🚫 Point Estimate Illusions — Mandatory confidence intervals and significance tests on all reported results.
  • 🧠 Hypothesis-Driven — Forces falsifiable goal-setting and a baseline expectation before touching raw data.

⚙️ How It Works: Skill Loading Flow

datapowers uses a hook-based on-demand loading architecture inspired by superpowers.

Session Start
     │
     ▼
hooks/session-start fires
     │
     ├── Reads skills/using-datapowers/SKILL.md
     │
     └── Injects it as session context (EXTREMELY_IMPORTANT)
               │
               ▼
         Agent now knows:
         • All 20 available skills
         • Trigger keyword for each skill
         • When to invoke each skill
               │
               ▼
         Analyst asks a question
               │
               ▼
         Agent matches trigger → invokes Skill tool
               │
               ├── Loads skills/<name>/SKILL.md (Iron Laws, Hard Gates, Procedure)
               │
               └── Loads supplementary files as needed:
                   • reference docs (e.g., assertion-anti-patterns.md)
                   • pressure tests (test-pressure-1.md, ...)
                   • agent prompt templates (analyst-prompt.md, ...)
                   • utility scripts (pipeline-pollution-detection.sh)

Key design principles:

  • using-datapowers is always loaded — it is the routing table for all other skills.
  • All other skills are loaded on demand — no unnecessary context inflation.
  • Supplementary files are scoped — each skill only loads what it needs.
  • Session state persists in artifacts/analysis_manifest.json — the agent can resume any session without losing context.

📋 Skills Reference

All 20 skills, organized by phase.

Phase 0 — Entry & State

SkillWhen to UseKey Output
using-datapowersAlways loaded at session startRouting table: all skills + trigger keywords
analysis-manifestSession start; "Where are we?"; after brainstormingartifacts/analysis_manifest.json — single source of truth for session state

Phase 1 — Design

SkillWhen to UseKey Output
brainstorming"Analyze X", "Design spec", "Hypothesis"docs/datapowers/specs/ design doc with ≥3 falsifiable hypotheses
writing-analysis-plansAfter brainstorming approvaldocs/datapowers/plans/ plan with 15–30 min tasks, no placeholders

Phase 2 — Data Understanding

SkillWhen to UseKey Output
data-profilingNew dataset; before any subagent dispatchartifacts/data_profile.md — high-density PII-free profile
data-explorationFirst-time dataset exploration; EDA requesteddocs/datapowers/eda/ report; leakage candidate flags
data-validationBefore any model training or feature engineeringPandera validation report; BLOCK / PROCEED verdict

Phase 3 — Feature Engineering & Modeling

SkillWhen to UseKey Output
leakage-guardTemporal dataset; before feature engineering reviewBLOCKED / NEEDS_HUMAN_REVIEW / APPROVED verdict
feature-engineeringBuilding or transforming featuresFitted transformers saved to artifacts/; Feature Registry updated
test-driven-data-scienceBefore any model.fit() callThree-layer assertion results; training blocked on CRITICAL failures
model-selectionChoosing between candidate modelsBaseline comparison table; Optuna HPO results (≥50 trials)
model-evaluationFinal test set evaluationBootstrap CIs; SHAP summary; one-time test set gate

Phase 4 — Execution & Review

SkillWhen to UseKey Output
executing-plans"Start tasks", "Follow plan"Two-stage review per task: Statistical → Code Quality
subagent-driven-analysisMulti-task parallel analysisParallel analyst subagents with isolated context; review gating
requesting-statistical-review"Audit results", "Significance", after task completionStatistical review verdict: APPROVED / ISSUES FOUND / BLOCKED
debugging-pipelinesPipeline error; unexpected model behavior; performance dropRoot cause investigation log; PSI drift report

Phase 5 — Delivery

SkillWhen to UseKey Output
verification-before-delivery"Done", "Complete", before any deliveryArtifact integrity checklist; reproducibility confirmation
report-writingFinal stakeholder reportReproducibility-header report with CIs and significance tests
finishing-an-analysis-branchAnalysis complete; ready to deliverCommit / PR / archive options; manifest-gated delivery

Meta

SkillWhen to UseKey Output
writing-data-skills"New skill", "Add skill", "Contribute skill"New SKILL.md passing Statistical Pressure Test (3 scenarios)

🚀 The Core Workflow

brainstorming → writing-analysis-plans
      │
      ▼
data-profiling → data-exploration → data-validation
      │
      ▼
leakage-guard → feature-engineering → test-driven-data-science
      │
      ▼
model-selection → model-evaluation
      │
      ▼
executing-plans (with requesting-statistical-review per task)
      │
      ▼
verification-before-delivery → report-writing → finishing-an-analysis-branch

At any step: analysis-manifest tracks completed stages. debugging-pipelines handles any unexpected failures. subagent-driven-analysis can parallelize independent tasks.


🛠️ Installation

Claude Code

# Install the plugin
/plugin install https://github.com/zpower426/datapowers

# Or clone and install locally
git clone https://github.com/zpower426/datapowers
cd datapowers
/plugin install .

How it works: .claude-plugin/plugin.json registers the hooks directory. On every session start, hooks/session-start fires and injects using-datapowers skill content into the session context automatically.

Gemini CLI

gemini extensions install https://github.com/zpower426/datapowers

The GEMINI.md file and gemini-extension.json manifest handle skill path registration.

OpenCode

git clone https://github.com/zpower426/datapowers ~/.opencode/plugins/datapowers

The .opencode/plugins/datapowers.js plugin injects bootstrap context via system prompt transform and auto-registers the skills directory.

Cursor / Codex

See .cursor-plugin/plugin.json and .codex/INSTALL.md for setup instructions.


⚖️ Iron Laws

DomainIron Law
EDANO MODELING WITHOUT EXPLORATORY DATA ANALYSIS
ValidationNO TRAINING WITHOUT DATA QUALITY VALIDATION
LeakageNO TRANSFORMERS FIT ON FULL DATASET BEFORE TRAIN/TEST SPLIT
EvaluationTEST SET EVALUATED EXACTLY ONCE — AT THE END
DeliveryNO DELIVERY WITHOUT CONFIDENCE INTERVALS AND SIGNIFICANCE TESTS
ReviewSTATISTICAL AUDIT ALWAYS PRECEDES CODE QUALITY REVIEW

📈 Star History

Star History Chart

🤝 Contributing

Contributions welcome. Read the writing-data-skills skill first — every new skill must pass the Statistical Pressure Test (balanced data, 1:100 imbalance, n < 200 small sample) before merge.

❤️ Acknowledgments

datapowers is inspired by and built on the shoulders of superpowers by Jesse Vincent (@obra).

The hook-based on-demand skill loading architecture, Iron Laws pattern, and Hard Gates design are all adapted from the superpowers family — extended here for the specific demands of statistical rigor in data science.

📜 License

MIT License

Ver o original no GitHub

Bom para

Relacionados