| name | ml-experimentation |
| description | Machine learning and data-science experiments — tabular, time series, or multi-table data. Use when the user explores a dataset, checks data quality, builds or trains a classifier/regressor, sets up sklearn or skrub pipelines, evaluates metrics, cross-validates, compares models, organizes an ML project, bootstraps a pixi/uv Python env for ML, iterates on experiments, audits results with skore, or asks in plain language: train a model, baseline pipeline, how good is this model, explore my data, feature engineering, predict X from Y, classification, regression, machine learning, tabular ML, skore report, next experiment. Students do not need to name Probabl skills — read the matching skill below and follow it.
|
ML experimentation (intent router)
Students describe goals in plain language. Do not ask them to pick a skill
name. Read the Probabl skill that matches their intent (under
~/.cursor/skills/ when probabl-skills is installed).
Route by intent
| User is trying to… | Read skill |
|---|
| Understand data, distributions, leakage, dtypes, missing values | explore-ml-data |
| Declare pipeline steps (load → preprocess → estimator), skrub DataOps | build-ml-pipeline |
| Fit/evaluate, metrics, plots, skore reports, CV | evaluate-ml-pipeline |
| Statistical + functional tests before shipping | test-ml-pipeline |
| Stress-test on future/unseen data shapes | smoke-test-ml-pipeline |
| Post-run audit from a skore report | audit-ml-pipeline |
| Track experiments, journal, design next run | iterate-ml-experiment |
| Diagnostics from skore → next experiment | iterate-from-skore |
| User proposes idea, paper URL, or reference repo | iterate-from-user |
| Project layout, cookiecutter-style workspace | organize-ml-workspace |
| Python style, typing, ruff-friendly patterns | python-code-style |
| pixi / uv / venv bootstrap for the project | python-env-manager |
| Which library for which job (pandas, polars, sklearn, …) | data-science-python-stack |
| Look up an installed package’s public API | python-api |
If several rows apply, follow pipeline order: explore → build → evaluate →
test → iterate. Do not skip explore when data is new.
AstroAI conventions
- Project env: pixi or uv under
${WORK} — not $HOME.
- Ray jobs:
astroai cluster start then astroai run script.py.
- Skore CLI (optional):
skore skills list · skore skills install — install
via astroai agent plugins install skore-cli.
Missing skills?
If Probabl skills are not installed:
astroai agent plugins install probabl-skills
That installs this router plus the full Probabl catalog.
