ds-crew

ds-crew — data science skills for Claude Code

See the whole arc: Walkthrough — from a raw file to a reviewed answer — a messy real dataset, a vague goal, and the plugin catching a silent revenue error on the way to a defensible number. (To record your own terminal cast, see docs/demo.md.)

14 slash commands that give Claude Code structured workflows for data science work — profiling, exploration, analysis, and model-building. The goal is a good kickoff: get from zero to a structured, executable, reviewed starting point in minutes rather than hours.

The contract. This is a kickoff tool, not an oracle.

• What it does: the mechanical work — profile the data, pin down the question, write and run the code, and check the result against common silent-failure modes.
• What it hands back to you: the judgment — the right question, the right data, the right interpretation — with the consequential scope decisions surfaced, not buried.
• When it stops and asks: ambiguous scope, contradictory requirements, or a domain judgment call. The cases it gets wrong are exactly the ones that warrant human or multi-agent review anyway — so it routes them to you instead of guessing.

→ Quickstart — one dataset, one command, expected output

---

What it does

Drop a data file in your project and use a slash command to kick off whatever you need:

• Start an analysis — /ds-conduct peeks at your data, asks a few questions, and assembles the right workflow automatically
• Profile a new dataset — /data-profile gives you column stats, null rates, and join diagnostics before you write a single line
• Explore without a question — /eda-narrative surfaces what's interesting and writes it up
• Get a verified answer — /ds-star-plus writes and runs Python iteratively, checks the result against 7 common failure modes (wrong column, dropped rows, unit mismatch, scope error, format mismatch, question substitution, missing output token), and backtracks if something's off
• Run multiple approaches — /ds-spike runs 3 diverse solvers in parallel and reconciles them; useful when you need confidence or when two single runs disagreed
• Build a model — /ds-model runs an AIDE-style solution tree with leakage and CV discipline

The skills are prompts — structured workflows that guide Claude through a task. They don't replace your judgment; they give you a faster, more consistent starting point than writing the prompt from scratch each time.

---

Installation

1. Add the marketplace (one-time per machine)

claude plugin marketplace add AdamKrysztopa/ds-crew

2. Install the plugin — choose your scope

# All your projects
claude plugin install ds-crew@ds-crew --scope user

# This project only (saved to .claude/settings.json)
claude plugin install ds-crew@ds-crew --scope project

# Local override, not committed
claude plugin install ds-crew@ds-crew --scope local

3. Use

/ds-conduct       # start here if you're not sure — inspects data, asks, routes
/ds-star-plus     # iterative solver with rubric-graded verification
/ds-spike         # 3 parallel solvers → consensus + minority report
/data-profile     # data quality / profiling report
/eda-narrative    # exploration → stakeholder narrative
/ds-star          # baseline solver (paper-faithful, single model)
/ds-clarify       # pin down intent before solving — writes analysis-spec.md
/ds-model         # AutoML solution-tree with leakage/CV discipline
/ds-verify        # grade any answer against the 7 DS failure modes
/ds-reconcile     # reconcile multiple existing answers into consensus
/ds-vote          # run same solver N times, take majority
/ds-search        # MCTS tree-search for a task that keeps failing greedy
/ds-memory        # remember and reuse recipes across sessions
/ds-env-setup     # verify Python env, install packages, offer SessionStart hook

First time? Run /ds-env-setup to verify your Python environment has the packages these skills need.

Updating / uninstalling

claude plugin update ds-crew@ds-crew
claude plugin uninstall ds-crew@ds-crew

After installing or updating: start a new Claude Code session (or /restart) — the plugin cache loads at session start.

---

The fourteen skills

Not sure where to start? Run /ds-conduct or see docs/USAGE.md for a decision table.

Core 5 — reach for these first