agent-ml-skills

Production-grade Machine Learning, Data Science & MLOps skills for AI coding agents.

Coding agents are great generalists but make the same ML mistakes over and over: leaking preprocessing into cross-validation, scoring imbalanced data with accuracy, forgetting model.eval(), building RAG with dense-only retrieval. agent-ml-skills is a curated pack of 15 battle-tested skills that teach your agent how an experienced ML engineer actually works — so it stops guessing.

Works with Codex, Claude Code, Cursor, and OpenCode.

Quick start

Install all skills into your agent with one command — no install, no dependencies:

# Codex
npx agent-ml-skills install --target codex

# Claude Code
npx agent-ml-skills install --target claude

# Cursor
npx agent-ml-skills install --target cursor --scope project

# OpenCode
npx agent-ml-skills install --target opencode

# Everything, everywhere
npx agent-ml-skills install --target all

Browse what's inside first:

npx agent-ml-skills list

Then restart your agent (or start a new session) and it will pick the right skill up automatically when your task matches.

What's a "skill"?

A skill is a single Markdown file with YAML frontmatter telling the agent when to use it and how to do the task well:

---
name: sklearn-pipelines
description: Use when building scikit-learn models that must not leak preprocessing...
---

# scikit-learn Pipelines
...workflow, code patterns, pitfalls, hand-off...

Agents that support skills load the description up front and pull in the full body only when the task matches — so you get expert guidance without bloating every prompt.

The skills

Data prep & exploration

| Skill | Use when… | |---|---| | exploratory-data-analysis | Starting on a new dataset — profiling, distributions, correlations, leakage & viz. | | data-cleaning | Handling missing values, duplicates, types, outliers — with train-only imputation. | | feature-engineering | Encoding, scaling, datetime/text/aggregation features, leakage-safe target encoding. | | pandas-patterns | Writing idiomatic, vectorized, memory-efficient pandas (no SettingWithCopyWarning). | | imbalanced-data | The target is rare (fraud/churn/disease) — metrics, SMOTE, class weights, thresholds. |

Modeling

| Skill | Use when… | |---|---| | sklearn-pipelines | Building scikit-learn models that must not leak preprocessing into CV. | | pytorch-training-loop | Writing/reviewing a PyTorch loop — eval modes, AMP, checkpointing, devices. | | model-evaluation | Choosing metrics, validating, calibration, confusion-matrix analysis. | | hyperparameter-tuning | Optimizing params — random vs Optuna, leakage-safe CV, early stopping, budget. |

LLMs & GenAI

| Skill | Use when… | |---|---| | llm-finetuning | Fine-tuning an LLM — full vs LoRA/QLoRA, data formatting, transformers/PEFT/TRL. | | rag-pipeline | Building RAG — chunking, embeddings, hybrid + reranking retrieval, eval. |

MLOps & reliability

| Skill | Use when… | |---|---| | experiment-tracking | Experiments need comparing/reproducing — MLflow/W&B, what to log, registry. | | reproducible-ml | A result must be reproducible — seeds, env pinning, data versioning, CUDA determinism. | | ml-debugging | A model won't learn, loss is NaN, or metrics look too good — a diagnosis decision tree. | | model-serving | Deploying behind an API — FastAPI, safe artifact loading, batching, ONNX, monitoring. |

Usage

npx agent-ml-skills <command> [options]

Commands
  list                       List available skills
  install                    Install skills into an agent

Options
  -t, --target <name>        codex | claude | opencode | cursor | all
      --scope <scope>        global (default) | project
      --skills <a,b,c>       comma-separated subset (default: all)
      --dir <path>           install into a custom directory (overrides target)
  -f, --force                overwrite existing skills
  -h, --help                 show this help

Examples

# Just the LLM skills, into the current project
npx agent-ml-skills install --target claude --skills rag-pipeline,llm-finetuning --scope project

# Into a custom agent directory
npx agent-ml-skills install --dir ./my-agent/skills

# Re-install and overwrite
npx agent-ml-skills install --target codex --force

Where skills get installed

| Target | Global | Project | |---|---|---| | Codex | ~/.codex/skills | .codex/skills | | Claude Code | ~/.claude/skills | .claude/skills | | OpenCode | ~/.config/opencode/skills | .opencode/skills | | Cursor | — | .cursor/rules (flat .md rules) |

Design principles

• Leakage-safe by default. Every data skill fits transforms on train only.
• Concrete over abstract. Real code patterns, not vague advice.
• Pitfalls included. Each skill ends with the mistakes agents actually make.
• Composable. Skills hand off to each other (EDA → cleaning → features → pipeline → eval → serving).
• Zero-dependency installer. Pure Node, nothing to install, nothing to trust.

Contributing

New skills and improvements are very welcome — see CONTRIBUTING.md. Every skill is validated in CI:

node scripts/validate-skills.mjs

Open a skill request if there's an ML workflow you want your agent to master.

Author

Built by Param Bhavsar — Google Summer of Code '19 @ TensorFlow, ex-HSBC. If this saves you a debugging session, a ⭐ helps others find it.

License

MIT