Datapowers
A complete data-science methodology for your coding agents, built on top of a set of composable skills and initial instructions that make sure your agent uses them.
Inspired by obra/superpowers, Datapowers adapts the same pattern (process skills first, Iron Laws, TDD-for-docs, subagent-driven execution) to the full ML lifecycle: framing, data, EDA, features, modeling, validation, deployment, monitoring, and incident response.
How it works
When you start a session in a project with Datapowers installed, your agent reads CLAUDE.md, loads the using-datapowers bootstrap skill, and from then on every relevant action triggers the right skill automatically:
- Ask the agent to "predict churn" → it invokes
framing-the-ml-problem and asks the right questions BEFORE touching any notebook.
- Agree on the spec → it invokes
writing-experiment-specs, commits a spec doc, then writing-experiment-plans to produce bite-sized tasks.
- Execution →
subagent-driven-experimentation dispatches one runner per configuration, with two-stage review (metrics-vs-spec, code-quality) after each.
- Before deploy →
reproducibility-verification, writing-model-cards, pre-deployment-review are gates that cannot be skipped.
- After deploy →
monitoring-data-and-model-drift and incident-response-for-ml keep watch.
The Iron Laws
Every critical skill defines an inviolable rule. The key ones for Datapowers:
- No model training without a baseline and a validation strategy declared first.
- No
fit() on the full dataset — never. Validation strategy is committed before any model is fit.
- No "model ready" claims without fresh reproducibility verification in a clean environment.
- No deployment without a model card, monitoring plan, and rollback plan.
- No production incident "fix" without root-cause investigation across data, model, and infra layers.
Lifecycle map
Bootstrap → Framing → Planning → Data → EDA → Features →
Modeling → Validation → Pre-Deploy → Deploy → Monitor & Ops →
Collaboration & Close
30 skills total, grouped in 11 phases. See skills/ for the full catalog.
Persistent artifacts
Specs, plans, model cards, and postmortems are committed to git in:
docs/datapowers/specs/YYYY-MM-DD-<topic>-spec.mddocs/datapowers/plans/YYYY-MM-DD-<feature>.mddocs/datapowers/model-cards/<model>-<version>.mddocs/datapowers/postmortems/YYYY-MM-DD-<incident>.md
Installation
This is a Claude Code / Cowork plugin. To use it locally, point your agent at this directory and ensure CLAUDE.md is loaded at session start.
Stack assumptions
Skills are written assuming a modern Python ML stack: Python 3.11+, pandas/polars, scikit-learn, PyTorch, HuggingFace, MLflow, DVC, pandera, evidently, BentoML / MLflow Models, FastAPI. Most patterns transfer to other stacks — substitute as needed.
License
MIT
