Probabl Skills

A set of skills to team up with you in your machine learning experimentation journey. It helps you at:

• organizing your workspace
• building your machine learning pipeline with the right libraries ensuring good methodologies
• evaluating and storing your results such that you can easily audit and get insights from them
• couple it with Skore Hub to get a comprehensive view of your experiments and their results
• iterate on your next experiments taking insights thanks to Skore diagnostics and your own feedback

So we aim at allowing you to focus on the science, letting AI agents to take care about the implementation but guided by two important ingredients: great libraries for the maintainability and good methodologies to make experiments right.

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 are working pretty well with models such as Claude Opus and Sonnet and gives really good results with smaller models such as Qwen 3.6 30B or DeepSeek v4 Flash. In terms of agent's harnessing, we tested them with Claude Code, OpenCode, Cursor, GitHub Copilot and do not witness any significant difference in terms of skills 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:

skore skills install

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 skills install

or

pixi exec --spec skore-cli skore skills install

If you prefer npx, then you can use:

npx skills add probabl-ai/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/skills

/plugin install probabl-skills@probabl-skills

/plugin update pulls new releases.

Skills in details

ML pipeline lifecycle

Skill	Description
explore-ml-data	Explore the dataset before designing any model.
build-ml-pipeline	Build a machine learning pipeline from the data source to the learner, including multi-tables engineering.
evaluate-ml-pipeline	Evaluate a complex machine learning pipeline and get structured reports including metrics, plots, and diagnostics.
test-ml-pipeline	Make sure that your machine learning pipeline is production-ready statistically and functionally.
smoke-test-ml-pipeline	Stress test your machine learning pipeline on future data to make sure it works.
audit-ml-pipeline	Once testing and the experiment is done, audit by loading a skore report and investigate.

Iteration loop

Skill	Description
iterate-ml-experiment	Design, keep track of experiments and iterate on them.
iterate-from-skore	Use skore to run diagnostics and checks that can be reported and addressed in the next experiment.
iterate-from-user	As a user be in the loop and propose new experiments — free-text, a scientific article URL, or a resource link (GitHub issue / spec / reference repo).

Workspace and tooling

Skill	Description
organize-ml-workspace	An organized workspace to keep track of your experiments.
python-code-style	Enforce good practices out-of-the-box for the Python ecosystem for your code.
python-env-manager	Bootstrapping the experiment setup based on your favorite Python environment manager.
data-science-python-stack	Opinionated one-library-per-job Python stack, organized into mandatory / user-choice / optional / transitive tiers.

API references

Skill	Description
python-api	Discover the public API of any installed Python package to make agent find their way without bothering your workspace.