Probabl Skills

A set of skills to steer your AI-assisted machine learning experiments. The skills help you:

• build your machine learning pipeline with core data science libraries (e.g. scikit-learn, skrub, skore, pandas, polars) while ensuring your agent follows correct methodologies
• evaluate and store your results so you can easily audit and get insights from them
• connect your agent to Skore Hub to get a comprehensive view of your experiments and results
• iterate on your next experiments from a Skore audit digest (a separate backlog turn) and from your own feedback
• organize your workspace according to best practices for data science projects (e.g. cookiecutter template)

Probabl skills let you focus on the science while AI agents handle the implementation, guided by two important ingredients: core data science libraries for maintainability and methodological best practices for running your machine learning experiments properly.

In practice, from a prompt such as:

╭────────────────────────────────────────────────────────────────────────╮
│ > Given the context in the file `data/README.md` and the data located  │
│   in `data/`, let's build a first machine learning pipeline that will  │
│   serve as baseline for the next experiments that we are going to run  │
│   together.                                                            │
╰────────────────────────────────────────────────────────────────────────╯

you can expect your agent to start experimenting with you. The skills work well with models such as Claude Opus and Sonnet and produce great results with smaller models such as Qwen 3.7 Flash or DeepSeek v4.1 Flash. Each skill declares metadata.modelTier (small, medium, or big). The eval maps those tiers to Qwen3.8 27B, DeepSeek v4.1 Flash, and Qwen3.8 Max. A skill may also declare metadata.role (entry or helper) when its caller continues after it loads. A helper runs on at most its caller's tier, so a skill never loads a helper of a higher tier; tools/validate_catalog.py enforces it.

As for agent harnesses, we tested them with Claude Code, OpenCode, Cursor, and GitHub Copilot and found no significant difference in terms of skill invocation.

Install

You can install the skills using the skore CLI that you can install from PyPI or from conda-forge and run the following command.

First install skore-cli:

# with pip
pip install skore-cli
# with uv
uv tool install skore-cli
# with pixi
pixi global install skore-cli

Then run the following command:

skore skills install

Install a smaller workflow pack by id when you do not need the full companion:

skore skills install setup  # workspace, environment, git, export
skore skills install data_analysis  # data exploration
skore skills install model  # frame, build, smoke-test, evaluate, audit
skore skills install loop   # triage, explore, model, review, backlog, export
skore skills install export  # notebooks and documentation site

skore skills install ml-experimentation remains the complete pack, containing every active skill. The default install / install all behavior is unchanged.

You can use uvx or pixi exec to install the skore CLI and directly run the command in an isolated environment:

uvx --from skore-cli skore skills install

or

pixi exec --spec skore-cli skore skills install

If you prefer npx, then you can use:

npx skills add probabl-ai/skore-skills

Alternative — Claude Code plugin marketplace

If you only use Claude Code and prefer the native plugin flow, this repo is also a Claude Code plugin marketplace:

/plugin marketplace add probabl-ai/skore-skills

/plugin install probabl-skills@probabl-skills

/plugin update pulls new releases.

Skills in detail

Meta and setup actions

ML pipeline lifecycle

Ideas and backlog

Workspace and tooling

Canonical package policy lives in the CLI; print it with python -m skore_skills env stack. choose-python-library resolves competing libraries.

When executed notebooks or the documentation site are enabled in a workspace, workflow stages publish them at completed checkpoints before the next user decision. Python percent files execute once: that run can write the notebook, optional notebook HTML, and audit digest together. Site checkpoints rebuild only when Markdown, notebook viewers, figures, or result viewers changed. Unset or disabled notebook/site policies do not enable these artifacts automatically.

Compatibility and removed skills

The catalog temporarily retains two deprecated compatibility skills. They are not included in workflow packs:

Catalog ids python-env-manager and python-code-style have been removed. Run skore skills remove on any leftover sidecars and reinstall the setup pack.

Website

The catalog page lives in site/ and is published at https://probabl-ai.github.io/skore-skills/.

Build it with the site Pixi environment, which provides Node.js:

pixi run -e site site-build

Pull requests that change site/, .catalog.json, or the Pixi manifest build the site in CI. GitHub Pages builds it from main when those same paths change. A repo admin turns this on once: Settings → Pages → Source GitHub Actions. If Pages is enabled after that workflow is already on main, run the Pages workflow with Run workflow. Catalog edits still go through pixi run check.