Research Program Skill

A portable research workflow skill for Claude Code, Codex CLI, and Cursor Agent.

It combines two useful ideas:

• huggingface/ml-intern: discovery, evidence labeling, and trace discipline for research work
• karpathy/autoresearch: baseline-first, metric-first, one experiment at a time

The default mode is one-shot. Run one baseline or one experiment, record the result, then stop.

Install for Claude Code

mkdir -p ~/.claude/skills
cp -R skills/research-program ~/.claude/skills/research-program

Then invoke:

/research-program

Use with Codex CLI or Cursor Agent

Copy examples/program.md into the target repo, then tell the agent:

Read program.md and follow it exactly. Run the baseline if none exists. Otherwise run one experiment only, record results.tsv, then stop.

When to use

Use this for:

• paper, repo, dataset, model, benchmark, or docs discovery
• turning a research topic into a reproducible experiment plan
• creating a shared program.md that multiple coding agents can follow
• running controlled baseline and experiment cycles
• adapting ml-intern or autoresearch ideas without giving those tools separate model API keys

Do not use it for unrestricted autonomous loops. Continuous or overnight mode should require explicit user instruction.

Repository layout

skills/research-program/SKILL.md  # Claude Code skill
examples/program.md               # reusable research program template
examples/results.tsv              # result log example

Safety model

The workflow keeps research bounded by requiring:

• a target metric before experiments
• a baseline before changes
• explicit allowed and forbidden files
• a results.tsv record for every run
• one run only unless the user explicitly asks for continuous mode

License

MIT