Skill: shared-git-data

Git for data science. Different from code — data needs its own rules.

Trigger

• Starting a data project and creating the repo
• Committing a notebook and you don't want to push megabytes of outputs
• Should the dataset be in the repo or not?
• You need to version pipelines without versioning heavy data

Workflow LEND

1. ANALYZE ├── Project: analysis, pipeline, dashboard? ├── Data: how big? can it go in the repo? (< 100 MB) ├── Notebooks: outputs included or stripped? └── Collaboration: solo or team?

2. OFFER (Senior Menu) ├── A) Standard git — code + clean notebooks + small data (< 100 MB) ├── B) Git + DVC — large data versioned outside the repo └── C) Git + nbstripout — notebooks without outputs, clean diffs

3. CHOOSE → user confirms

4. EXECUTE ├── .gitignore: .venv/, pycache/, *.pyc, .env, data/raw/ ├── nbstripout: git config filter.nbstripout.extrakeys 'metadata.kernel' ├── DVC: dvc init && dvc add data/raw/dataset.csv && git add data/raw/dataset.csv.dvc ├── README.md with reproduction instructions ├── requirements.txt frozen ├── Commits in English US, semantic (feat, fix, chore) └── Branch + PR workflow for data pipelines

5. VERIFY ├── .gitignore covers what shouldn't be in the repo ├── Notebooks can be diffed (no outputs) └── A fresh git clone && pip install is enough to start

• Save decisions and changes to Engram (mem_save)
• Check if README, AGENTS.md, or ARCHITECTURE.md need updating
• If docs changed → update them in the same PR/commit