DataPowers
zpower426/datapowers
Protocolo de auditoria estatística para pipelines de dados robusto e à prova de vazamentos
Ver no GitHubProtocolo 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/datapowers1 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
- 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-datapowersis 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
| Skill | When to Use | Key Output |
|---|---|---|
using-datapowers | Always loaded at session start | Routing table: all skills + trigger keywords |
analysis-manifest | Session start; "Where are we?"; after brainstorming | artifacts/analysis_manifest.json — single source of truth for session state |
Phase 1 — Design
| Skill | When to Use | Key Output |
|---|---|---|
brainstorming | "Analyze X", "Design spec", "Hypothesis" | docs/datapowers/specs/ design doc with ≥3 falsifiable hypotheses |
writing-analysis-plans | After brainstorming approval | docs/datapowers/plans/ plan with 15–30 min tasks, no placeholders |
Phase 2 — Data Understanding
| Skill | When to Use | Key Output |
|---|---|---|
data-profiling | New dataset; before any subagent dispatch | artifacts/data_profile.md — high-density PII-free profile |
data-exploration | First-time dataset exploration; EDA requested | docs/datapowers/eda/ report; leakage candidate flags |
data-validation | Before any model training or feature engineering | Pandera validation report; BLOCK / PROCEED verdict |
Phase 3 — Feature Engineering & Modeling
| Skill | When to Use | Key Output |
|---|---|---|
leakage-guard | Temporal dataset; before feature engineering review | BLOCKED / NEEDS_HUMAN_REVIEW / APPROVED verdict |
feature-engineering | Building or transforming features | Fitted transformers saved to artifacts/; Feature Registry updated |
test-driven-data-science | Before any model.fit() call | Three-layer assertion results; training blocked on CRITICAL failures |
model-selection | Choosing between candidate models | Baseline comparison table; Optuna HPO results (≥50 trials) |
model-evaluation | Final test set evaluation | Bootstrap CIs; SHAP summary; one-time test set gate |
Phase 4 — Execution & Review
| Skill | When to Use | Key Output |
|---|---|---|
executing-plans | "Start tasks", "Follow plan" | Two-stage review per task: Statistical → Code Quality |
subagent-driven-analysis | Multi-task parallel analysis | Parallel analyst subagents with isolated context; review gating |
requesting-statistical-review | "Audit results", "Significance", after task completion | Statistical review verdict: APPROVED / ISSUES FOUND / BLOCKED |
debugging-pipelines | Pipeline error; unexpected model behavior; performance drop | Root cause investigation log; PSI drift report |
Phase 5 — Delivery
| Skill | When to Use | Key Output |
|---|---|---|
verification-before-delivery | "Done", "Complete", before any delivery | Artifact integrity checklist; reproducibility confirmation |
report-writing | Final stakeholder report | Reproducibility-header report with CIs and significance tests |
finishing-an-analysis-branch | Analysis complete; ready to deliver | Commit / PR / archive options; manifest-gated delivery |
Meta
| Skill | When to Use | Key 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
| Domain | Iron Law |
|---|---|
| EDA | NO MODELING WITHOUT EXPLORATORY DATA ANALYSIS |
| Validation | NO TRAINING WITHOUT DATA QUALITY VALIDATION |
| Leakage | NO TRANSFORMERS FIT ON FULL DATASET BEFORE TRAIN/TEST SPLIT |
| Evaluation | TEST SET EVALUATED EXACTLY ONCE — AT THE END |
| Delivery | NO DELIVERY WITHOUT CONFIDENCE INTERVALS AND SIGNIFICANCE TESTS |
| Review | STATISTICAL AUDIT ALWAYS PRECEDES CODE QUALITY REVIEW |
📈 Star History
🤝 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.