Set Up Workspace

Decide where artifacts live. Do not design an experiment here.

Human-facing prose

Details: references/human_facing_prose.md. Scaffolded design notes, JOURNAL, and folder READMEs describe this workspace's analysis — not the skills framework, the CLI, or the command that produced an output. Questions and replies use the same data-science language (import name, notebooks, documentation site) — not G-* names, catalog skill ids, or the wrapper CLI. Say "workspace setup", not setup-workspace. <!-- results-embed: … --> is a site marker. Authoring hints stay in this skill. style is ruff only.

Detection

• src/ or journal/ present → existing. Reuse names and folders. Do not scaffold again.
• Manager manifest without src/<pkg>/ → manager-only. Still scaffold; keep existing pyproject.toml. No --force.
• Otherwise → fresh. Ask G-PKG-NAME, then scaffold.

Lookup

Detect the layout, then take the one row that matches and run its steps from the Sequence below. Do not stop after reading the table.

An unresolved import name is the ask row only. Steps 2, 4, and 5 run on a later turn, once policy.package or src/<pkg>/ names it. Step 4 does not overwrite a flag that is already true or false. Step 5: a turn dispatched by setup-ml-project returns to it; a standalone turn runs git end-turn --stage setup.

Sequence

1. If fresh or manager-only, resolve G-PKG-NAME. When it is not already recorded, call AskUserQuestion and stop this turn. Do not persist policy.package and do not scaffold until a later turn. “You pick” / “go fast” / “no preference” does not record it. The question is the src/<pkg>/ import name, with the folder name as the default option. Each ask states in 2–4 lines what the answer authorizes — the scaffolded tree, the persisted policy key, the toolchain a gate implies — and the detected facts it rests on; a file link is an addition, never the context. Do not confirm in prose instead of the tool. A matching [project] name + src/<pkg>/ already resolves it — do not re-ask. A recorded policy.package also resolves it. Do not re-ask. When setup-ml-project dispatched this turn and policy.package is set, do not ask again. Scaffold is step 2, and only once the name is already resolved. A status package of null does not reopen the ask. Persist a resolved import name with python -m skore_skills policy set package <pkg>.

2. Fresh / manager-only:

   python -m skore_skills scaffold --package <pkg>

   The CLI writes the tree and each folder README.md. Do not recreate those files from memory.

3. Existing: do not scaffold again. Do not invent files. No rename, no overwrite, no --force. Do not persist or ask notebooks/site.

4. After scaffold on fresh or manager-only, if policy.notebooks and policy.site are both null (never set): persist both on this turn —

   python -m skore_skills policy set notebooks true

   python -m skore_skills policy set site true

   Do not AskUserQuestion for notebooks or site. Persist both before the user-facing line. That line says executed notebooks and the documentation site are on; the user can turn either off later. Do not say they are still unset or null. Do not write notebooks or site into JOURNAL; policy is the record. Do not ask.

   Same turn after persist:

   Load add-python-package only if status.skills.add-python-package is true and policy.env.managed is true. Else one-line skip: name jupytext / nbclient / ipywidgets / nbconvert / mkdocs-material; do not invent pixi add / uv add; skip site init. Do not invent that skill's steps.

   When that load is allowed:

   • notebooks true → add-python-package for jupytext, nbclient, and ipywidgets (all env route agent), plus nbconvert when site is also true — stage turns write the notebook viewer with --html. Do not leave them as ask. Do not convert.
   • site true → add-python-package for mkdocs-material (agent), then python -m skore_skills site init.

   If either flag is already true or false, do not overwrite and do not install from this step.

5. If setup-ml-project dispatched this turn and is in this session, return to it; else stop. Standalone: python -m skore_skills git end-turn --stage setup. If JSON action is invoke, load persist-ml-git only if status.skills.persist-ml-git is true and stop; it returns to triage. If persist is missing, name the pending staged paths and stop. Otherwise load triage-ml-task only if status.skills.triage-ml-task is true; else stop.

Stop conditions

• "You pick", "go fast", and "no preference" do not resolve the import name. Call AskUserQuestion and stop. Do not scaffold or persist in that turn, and do not use the folder name yourself.
• Do not write experiments/01_baseline.py or any experiment body. "So we can run it right away" is not a reason. Scaffold writes experiments/README.md only.
• Do not re-run skore.evaluate or project.put, and do not paste that snippet. A KeyError from project.get is not a scaffold turn. Say evaluation and project.put are not this skill, and stop.
• Do not ask env manager, tabular library, or skore mode.
• Do not run pixi init / uv init.
• Do not env-bootstrap or editable-install. Export toolchain (jupytext, nbclient, ipywidgets, nbconvert, mkdocs-material) only via add-python-package after persisting notebooks/site. Never pixi add / uv add from this skill.
• Do not write experiment or exploratory data analysis bodies.
• Never git commit here.
• Do not AskUserQuestion for notebooks or site.
• Do not persist notebooks/site on an existing layout.

Layout (CLI writes this)

src/<pkg>/
experiments/
journal/
data_analysis/
data/
audit/
tests/smoke/
scratch/

Each directory has README.md.