ml-competition-with-claude-template

An open-source workspace for reproducible machine-learning competitions. It combines leak-aware validation, experiment tracking, submission checks, and Claude Code guidance in one template for Kaggle, SIGNATE, atmaCup, Nishika, ProbSpace, and similar platforms.

日本語ドキュメント (Japanese README)

Why this template?

• Reproducible by default: lock CV decisions before modeling and record every experiment.
• Safe with competition data: raw and external data are immutable and ignored by Git.
• Claude Code-native: /setup, /eda, /baseline, and platform skills provide a consistent workflow.
• Free tooling first: uses uv and supports local CPU/CUDA/MPS plus free notebook runtimes.
• Quality gates included: Ruff, mypy, pytest, secret detection, and submission validation.

Quick start

Initialize the workspace with Claude Code:

The setup command populates .references/, adapts CLAUDE.md, and prepares the competition-specific structure for Claude Code.

Optional dependency groups:

Install all bundled agent skills into an existing competition project:

The installer updates only the bundled .claude/skills and .claude/commands paths. It does not delete unrelated project files or locally added skills.