Declare the pipeline from data source to predictor as a **skrub DataOps graph** (not as a bare `sklearn.Pipeline`). Every step is either a pure-Python function (stateless) attached via `.skb.apply_func`, or a sklearn-compatible estimator (stateful) attached via `.skb.apply`. Stops at the declared object — no fit, split, tuning, persistence, or evaluation. TRIGGER — any of: - Writing or editing code that declares any link in the chain *data source → predictor*: loaders, preprocessing, encoders / imputers / scalers, feature steps, composition objects (`Pipeline`, `ColumnTransformer`, skrub `tabular_pipeline`, `nn.Module`), or the final estimator. - A pure-Python data-processing function destined for the pipeline path (cleans / derives / reshapes) — whether wrapped via `FunctionTransformer`, `skrub.@deferred` / `skrub.var`, a custom `BaseEstimator` subclass, or just called in the training path before the estimator. - A step is added, removed, swapped, or reordered inside an existing pipeline de

2026-06-2370

evaluate-ml-pipeline

probabl-ai/skills

Methodology for evaluating a single sklearn-compatible learner (in particular, the `SkrubLearner` produced by `build-ml-pipeline`). Owns: which entry point to call (`skore.evaluate` first, the explicit report classes when needed), which cross-validator to pick from scikit-learn's catalogue, how to consume the structural metadata (`groups`, `times`, …) attached at build time via `.skb.mark_as_X(split_kwargs=...)`. Stops at "what does the report say". Defaults (metrics, plots) come from skore; only override on explicit user request. TRIGGER when: code calls `cross_val_score`, `cross_validate`, `classification_report`, or any handwritten metric print (`print(mean_squared_error(...))`); code calls `.skb.cross_validate(...)` (route through skore for richer output); user asks how to score, evaluate, or compare a single learner; user asks how to pick a cross-validator; user wants to see a report / metrics / diagnostic plots for a fitted learner. SKIP when: declaring the pipeline (use `build-ml-pipeline`); hyperparam

2026-06-2370

python-api

probabl-ai/skills

Look up the public API of a Python package against the *installed version* and cache what's worth keeping. Four shapes by question type: (0) cache hit under `scratch/api/<lib>/<version>/`; (1) `inspect.signature` + `pydoc.render_doc` for a symbol; (2) `dir` / `pkgutil.iter_modules` for a module surface; (3) WebSearch + WebFetch of versioned docs for narrative ("how", "which", "what does X return when Y"). Never write a symbol from training-data memory — recognition is not a lookup. TRIGGER — any of: - About to name a symbol (function / class / method / arg) in code. - User asks "what's the signature of X?", "what's in module Y?", "how do I call X?", "which of A/B should I use?". - User asks "what does X return when <condition>?" (Shape 3 — see decision table). - Another workflow skill (`build-ml-pipeline`, `evaluate-ml-pipeline`, `iterate-from-skore`, `smoke-test-ml-pipeline`) says "consult the API skill". - About to reach for a library's "obvious" pattern from memory. SKIP when: the signature is obvi

2026-06-2370

data-science-python-stack

probabl-ai/skills

Opinionated Python stack for data-science / ML work — one library per job, organized into tiers (mandatory / user choice / optional / transitive). SKILL.md is the index; per-library `references/<library>.md` files carry scope, "pick this when" / "pick something else when", and pairings. TRIGGER when (any of these): (1) **a library import fails** in this stack's domain — the answer is install, not substitute (see § "Missing dependency"); (2) **a library choice has to be made** — explicitly (the user asks "which library for X?") or implicitly (code is about to introduce a new dependency, or the project is being scaffolded and the tabular library hasn't been picked yet); (3) starting a new Python data-science / ML project; (4) the user or current code reaches for a substitute outside the stack (xgboost, lightgbm, black, isort, flake8, poetry, hatch), or reaches for `mlflow` to log params/metrics, or for `cross_val_score` + handwritten reporting — redirect: tracking → `skore` Project API, evaluation / reporting →

2026-06-2370

organize-ml-workspace

probabl-ai/skills

Decide where files live in an ML experimentation project: reusable code in `src/<pkg>/`, one `# %%` script per experiment in `experiments/`, design notes + index in `journal/`, reports in `reports/`, agent-only probes in `scratch/`. Owns the layout, the file-creation rules (one file per experiment, ask before editing), and the jupytext `# %%` script convention. Never imposes `data/` — the user owns that. TRIGGER — any of: - Starting a new ML project / scaffolding a workspace. - About to create the first experiment file in a project. - About to create `src/<pkg>/data.py` / `features.py` / `pipeline.py` / `evaluate.py` for the first time. - About to write a `.ipynb` for experimentation — redirect to a `# %%` script under `experiments/`. - User asks where something should live, how to organize the project, or how to set up the workspace. - About to add a new experiment iteration — decide new file vs edit existing (ask the user). SKIP when: the file is clearly part of an already-populated module (e.g., a

2026-06-2370

audit-ml-pipeline

probabl-ai/skills

Owns the `audit/` folder: one `# %%` (jupytext percent) Python file per experiment, aligned 1:1 with `experiments/NN_<short_name>.py` and `journal/NN_<short_name>.md`, that loads the experiment's skore report **read-only** and uses bare-last-expression cells whose `__repr__` carries the audit's signal. The agent executes the audit file via the bundled in-process runner (`audit-ml-pipeline/scripts/run_cells.py` — IPython `InteractiveShell.run_cell`), which streams a markdown digest of each cell's stdout + last-expression repr to stdout (optionally also to a file). The digest fuels narrative work (the `JOURNAL.md` Status + History update, follow-up questions about a past experiment, cross-experiment comparison). Stops at "audit/NN_*.py is placed, executed, and the digest is available." Never calls `skore.evaluate(...)` or `project.put(...)`. TRIGGER — any of: - `iterate-ml-experiment` § 4 record-outcome — audit is dispatched FIRST (replaces scratch probes for metric extraction). - The user asks "audit exper

2026-06-1370

name

test-ml-pipeline

description

Owns the `tests/` folder of an ML workspace and the pairing rule between an experiment and its tests. Lightweight router: every test category has its own subskill (`smoke-test-ml-pipeline` is the only one for v1; `regression-test-ml-pipeline` / `distribution-test-ml-pipeline` etc. plug in as siblings as the workspace grows). This skill places an empty `tests/<category>/` folder, enforces the stem-pairing rule between `tests/<category>/test_NN_<short_name>.py` and `experiments/NN_<short_name>.py`, and dispatches to the matching subskill when the user asks for a test. TRIGGER when: a new design note was just approved by `iterate-ml-experiment` and the matching test has to be drafted before the experiment can be marked done; the user asks "write the smoke test for `02`", "add the regression test", "do we have a test for X?"; an experiment script was edited and the paired test needs revisiting; about to run `pytest tests/` and one of the expected paired tests is missing. SKIP when: the design note does not yet exist (route to `iterate-ml-experiment` first); the test is for the package's internal helpers and unrelated to a specific experiment (regular unit tests live wherever the project's `pytest` config picks them up — out of scope here); the question is *what does the test result mean* rather than *should the test exist* (route to the matching subskill). HOW TO USE: this skill is a router. Read the dispatch table to figure out which subskill owns the test category the user is asking about, then hand off. Do not write the test body yourself — that belongs to the subskill. Do place the empty test file with the matching stem and the pytest scaffolding, then hand control over.

Test ML Pipeline (router)

Where tests for an ML workspace live, what gets paired with what, and which subskill owns the body of each test category.

First action (every turn)

Before answering anything else:

Confirm an approved design note exists for the test the user is asking about. The pairing rule is hard: tests/<category>/test_NN_<short_name>.py only exists if journal/NN_<short_name>.md is at least approved and experiments/NN_<short_name>.py is the matching script. If the design note doesn't exist, hand back to iterate-ml-experiment.
Emit the Pre-flight checklist (below) as visible text in your response, with each box marked.
Use the Dispatch table to pick the subskill that owns the test category, then hand off.

Pre-flight — emit this checklist as visible text before any test work

Pre-flight (test-ml-pipeline):
- [ ] `journal/NN_<short_name>.md` exists and is at least `approved`
      (or confirmed n/a — about to hand off to `iterate-ml-experiment`)
- [ ] `experiments/NN_<short_name>.py` exists with the matching stem
      (or confirmed n/a — about to hand off to `organize-ml-workspace`)
- [ ] Test category picked: smoke | regression | distribution | …
- [ ] Subskill dispatched: `smoke-test-ml-pipeline` | …
- [ ] Test file stem decided: `tests/<category>/test_NN_<short_name>.py`
- [ ] pytest is on the project's dependency manifest (per
      `data-science-python-stack` § Tier 1)

Stop conditions — read before anything else

No test without an approved design note. Never create tests/<category>/test_NN_*.py if the matching journal/NN_*.md isn't on disk and at least approved. The design note is the contract; the test asserts the contract holds. Reverse order is incoherent.
The stem rule is hard. Test file basename is test_NN_<short_name>.py (with the test_ prefix that pytest expects); the NN_<short_name> portion matches the experiment exactly. One experiment → one test file per category. No test_<NN>_v2.py, no test_NN_<short_name>_2.py. If a test needs to evolve, edit it in place; the pairing must stay 1:1.
One subskill per category. This skill only places the empty test file and hands off. Don't write assertion bodies, fixture construction, or test-specific logic in this skill — that belongs to the matching subskill (smoke-test-ml-pipeline, etc.).
pytest is the runner. Tests are pytest tests, not jupytext-style # %% scripts. The experiment scripts live in experiments/ and stay # %%-style for interactive iteration; the tests are binary pass/fail and benefit from pytest's reporting. Don't mix the two conventions.

Layout this skill owns

project/
└── tests/
    ├── smoke/
    │   ├── test_01_baseline.py        # ↔ experiments/01_baseline.py
    │   ├── test_02_short_name.py
    │   └── ...
    └── (future: regression/, distribution/, …)

The pairing rule:

journal/NN_<short_name>.md
experiments/NN_<short_name>.py
tests/<category>/test_NN_<short_name>.py

— same NN_<short_name> stem in all three. The test_ prefix on the test file basename is the pytest naming convention; everything after it tracks the experiment.

The tests/<category>/ subfolder lets the workspace grow more test types without renaming anything. tests/smoke/ is the only required category for v1.

Test categories — Dispatch table

Use the user's signal first; fall back to the defaults at the bottom.

Situation	Subskill
Brand-new experiment was just approved; need to wire its smoke test	`smoke-test-ml-pipeline`
User says "write the smoke test for `02`", "the smoke test is failing", "what should the smoke test for X assert?"	`smoke-test-ml-pipeline`
User says "the metrics drifted between `02` and `03`, can we lock that in?", "regression test against last week's run"	`regression-test-ml-pipeline` (future — not implemented in v1)
User says "predictions are out of range", "calibration looks off in production", "schema invariants"	`distribution-test-ml-pipeline` (future — not implemented in v1)
User asks an open-ended "should this experiment have a test?"	Default to `smoke-test-ml-pipeline` — the smoke test is required at every iteration; everything else is opt-in.

The current implementation only ships smoke-test-ml-pipeline. Future categories will land as sibling subskills; the dispatch table is the contract for adding them.

The required-test-per-experiment rule

Every approved experiment must have a passing smoke test before it can be marked done in JOURNAL.md. This is enforced by iterate-ml-experiment § 3 (after design-note approval, before script creation) and § 4 (before recording outcome): the test is part of the iteration loop, not an afterthought. If the smoke test fails, the iteration that follows is on the pipeline (re-enter build-ml-pipeline), not on the model — a smoke-test failure is almost always a structural problem with how the DataOps graph is laid out, not a metric problem with the predictor.

The smoke test is non-optional for every experiment in the workspace. Other test categories (regression, distribution, …) are opt-in and added when the workspace's needs warrant them.

Decision flow

Is there an approved journal/NN_<short_name>.md and a matching experiments/NN_<short_name>.py?
- No → hand off to iterate-ml-experiment (design note first) or organize-ml-workspace (script first). Stop.
- Yes → continue.
Identify the test category from the user's signal (Dispatch table). Default: smoke.
Place the empty test file at tests/<category>/test_NN_<short_name>.py with the pytest scaffolding (one def test_*(): function, empty body, a # TODO: filled in by <subskill> marker). If the file already exists, do not overwrite.
Hand off to the subskill (smoke-test-ml-pipeline etc.). The subskill writes the assertions, fixture construction, and any helpers it needs.
Confirm with the user that the subskill's draft is correct before running the test (the subskill owns this loop; this skill only does placement + dispatch).

What this skill does NOT do

Run pytest. Test execution is the user's call (or CI's).
Write assertion bodies. Each subskill owns the assertions for its category.
Decide whether a test is required. The required-test-per-experiment rule (smoke at every iteration) is fixed by this skill; adding optional categories is a workspace-level decision the user makes, not this skill.
Touch files outside tests/. The companion-skill edits to iterate-ml-experiment / organize-ml-workspace / build-ml-pipeline / evaluate-ml-pipeline are on those skills, not this one.

Companion skills

smoke-test-ml-pipeline — owns the smoke test contract (fixture construction, the diagnostic-by-construction property, assertions, failure semantics). The only required test category in v1.
organize-ml-workspace — scaffolds tests/<category>/ alongside journal/ / experiments/ at workspace creation time. The placeholder test files are placed by this skill, not by organize-ml-workspace.
iterate-ml-experiment — drives the iteration loop. After design-note approval, dispatches to this skill (which dispatches to smoke-test-ml-pipeline) to draft the matching test. After the experiment runs, requires the smoke test to pass before the experiment can flip to done.
build-ml-pipeline — pipeline declaration. The smoke test is the executable proof that build-ml-pipeline's X-marker rule (mark X early, featurize after, history references via upstream nodes) was followed. Smoke test failure typically means the pipeline shape is wrong; route back here.
evaluate-ml-pipeline — owns the CV protocol. CV is necessary but not sufficient when the pipeline has history-dependent features; smoke-test-ml-pipeline fills the gap CV doesn't.
data-science-python-stack — declares pytest as a Tier 1 mandatory dependency for any workspace that uses this skill.