data-science-setup

The saved excerpt is a snapshot from review. The external source remains the complete and most current version.

---
name: data-science-setup
description: Set up Claude for data science and ML engineering — configures notebook workflow, experiment tracking, reproducibility conventions, and data layout so Claude supports you from exploration to productive modeling.
---

# Data Science / ML Setup

This skill configures Claude for exploratory and productive data science / ML work. It is the right choice when your project centers on notebooks, models, datasets, and experiments — not general application code (use `coding-setup` for that) and not literature research (use `research-setup` for that).

**Handoff context:** Read `skills/_shared/consume-handoff.md` and run it with the handoff block (if any). The helper guarantees the following locals: `detected_language`, `existing_claude_md`, `inferred_use_case`, `repo_signals`, `graphify_candidate`. Use `detected_language` for all user-facing prose; generated file content stays in English.

**Existing CLAUDE.md:** If `existing_claude_md: true`, DO NOT overwrite it. Append a new delimited section at the end of the file:

```
<!-- onboarding-agent:start setup=data-science skill=data-science-setup section=claude-md -->
## Claude Onboarding Agent — Data Science Setup
...generated content...
<!-- onboarding-agent:end -->
```

If the delimited block already exists from a previous run (either the attributed form above or the legacy unattributed `<!-- onboarding-agent:start -->` form), replace only the content between the markers; leave the rest of the file untouched. Upgrade the opening marker to the attributed form while you are there — `/upgrade-setup` depends on it for detection.

## Supporting Files

Read these on-demand at the step that invokes them. Do not read eagerly.

- `rule-file-templates.md` — bodies of the `.claude/rules/*.md` files (Step 4)
- `stack-scaffolds.md` — `pyproject.toml`, `uv add` commands, `.claude/settings.json` permissions, directory scaffold (Step 4)
- `gitignore-block.md` — the `.gitignore` block (Step 4)
- `notebook-hygiene.md` — `.pre-commit-config.yaml` for nbstripout + nbqa (Step 4)
- `skills/_shared/consume-handoff.md` — orchestrator handoff parse + inline fallback (preamble, before Step 1)
- `skills/_shared/offer-superpowers.md` — canonical Superpowers opt-in (Step 1)
- `skills/_shared/offer-graphify.md` — canonical Graphify opt-in (Step 6)

## Step 1: Install Dependencies

Read `skills/_shared/offer-superpowers.md` and run it with `skill_slug: data-science-setup`, `mandatory: false`, `capability_line: "A widely used free Claude Code skills library. Useful for planning multi-step experiments and structuring model-training pipelines."` The helper asks the user, delegates to `skills/_shared/installation-protocol.md` on `yes`, and sets `superpowers_installed`, `superpowers_scope`, `superpowers_method`.

## Step 2: Verify Python Tooling

Run `uv --version` (via Bash).

- If the command succeeds: set `uv_available: true` and continue.
- If it fails: set `uv_available: false` and print ONCE:

  > "⚠ `uv` is not installed. This skill strongly recommends `uv` as the Python package manager for data-science projects (reproducible lockfiles, fast installs, no global pollution). Install it from https://docs.astral.sh/uv/getting-started/installation/ and re-run this skill when ready. Setup will continue, but `pyproject.toml` and `uv add` steps will be emitted as instructions only — nothing will be executed on your machine."

Never try to install `uv` automatically, and never fall back to `pip`, `poetry`, or `conda`.

## Step 3: Context Questions

Ask these questions ONE AT A TIME. Wait for each answer before asking the next.

1. "Which language / stack is this project built on?
   A) Python + uv (default, recommended)
   B) R
   C) Julia
   D) Mixed (Python + one of the above)"

2. "Which notebook tool do you use?
   A) Jupyter (classic / JupyterLab)
   B) marimo (reactive, git-friendly)
   C) VS Code notebooks
   D) None / plain scripts only"

3. "Which deep-learning / heavy ML f