ml-agent-skills

a collection of machine learning skills for ai agents. this repository provides structured instructions (SKILL.md files) that enable ai agents to perform end-to-end ml workflows correctly.

philosophy

"teach the agent, don't script it."

• SKILL.md files contain senior data scientist knowledge: best practices, code patterns, and decision rules
• agents write their own code following the instructions
• flexible and adaptable to any codebase or context

quick start

# clone the repository
git clone https://github.com/your-username/ml-agent-skills.git
cd ml-agent-skills

# install python dependencies
pip install pandas numpy scikit-learn xgboost matplotlib seaborn joblib

skills overview

skill	purpose	what it teaches
ml-eda-viz	exploratory data analysis	distributions, correlations, leakage detection
ml-data-prep	data cleaning & splitting	stratified splits, imputation, encoding
ml-train-tabular	model training	pipelines, cross-validation, early stopping
ml-evaluate	model evaluation	metrics selection, threshold tuning, diagnostics

recommended workflow

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│  ml-eda-viz │ ──▶ │ml-data-prep │ ──▶ │ml-train-    │ ──▶ │ ml-evaluate │
│  (explore)  │     │  (clean &   │     │  tabular    │     │   (test)    │
│             │     │   split)    │     │  (train)    │     │             │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘

for ai agents

if you're an ai agent (claude code, cursor, etc.), read the AGENTS.md file for:

• workflow order and trigger phrases
• critical rules to prevent data leakage
• output conventions

each skill folder contains a SKILL.md with:

• detailed best practices and code patterns
• example code you can adapt
• common pitfalls to avoid
• checklists to verify correct implementation

best practices embedded

this repository encodes senior data scientist knowledge:

• data leakage prevention: split before computing any statistics
• stratified splitting: preserve class distributions (default 70/15/15)
• cross-validation: use stratified k-fold within training
• pipelines: preprocessing inside cv, not before
• early stopping: prevent overfitting in boosting models
• proper evaluation: never evaluate on training data; use f1/roc-auc for imbalanced datasets
• reproducibility: always set random_state

requirements

• python 3.10+
• pandas, numpy, scikit-learn, xgboost, matplotlib, seaborn, joblib

license

MIT