ai-co-scientist_skill

Overview

This skill turns Claude Code into an AI Co-Scientist that runs reproducible, hypothesis-driven research using tree-based exploration. It orchestrates project initialization, staged workflows, systematic experimentation, and automated git commits to capture provenance. The skill is domain-agnostic and enforces user checkpoints before each stage.

How this skill works

I initialize a project directory with state and visualization files, then guide research through five stages: literature review, hypothesis formulation, experimental design, systematic experimentation, and validation. During experimentation I use a tree search to propose, add, run, and evaluate experiment nodes, and I require git commits after each stage and experiment to ensure reproducibility. At every stage I prompt you to verify variables, budgets, and plans before proceeding.

When to use it

• Starting a new computational research or data-driven science project
• Structuring reproducible experiments that need systematic exploration
• Managing multi-step studies that require checkpoints and versioned results
• Exploring large hypothesis spaces where best-first or tree search helps prioritize runs
• Preparing experiment provenance for papers or audits

Best practices

• Define a clear, falsifiable hypothesis before designing experiments
• Confirm independent, dependent, and control variables at each checkpoint
• Set a realistic compute and iteration budget up front to bound the tree search
• Commit work frequently with descriptive messages that summarize findings and next steps
• Validate promising configurations with multiple seeds and ablations before final synthesis

Example use cases

• Test whether aggressive data augmentation improves model robustness across augmentation intensities
• Compare optimization algorithms across hyperparameter branches using tree-based expansion
• Explore model architecture variants where promising branches are expanded and less promising ones pruned
• Run ablation studies and validation runs after identifying top-performing configurations
• Collect reproducible experiment artifacts and generate figures for a methods section