How to use the Experiment Bridge agent skill

What is an agent skill, and what does experiment-bridge do?

An agent skill is a folder containing a SKILL.md file. The file tells an agent like Claude Code when the skill applies and what steps to follow. This skill is named experiment-bridge and is described as Workflow 1.5 in the Auto-claude-code-research-in-sleep project.

When it is active, the agent reads EXPERIMENT_PLAN.md, produced by an earlier research step, and writes the experiment code. That covers training, evaluation, baselines, fixed seeds and JSON or CSV result files. It can optionally ask a second AI model to review the code for bugs, then deploy the runs to a GPU and collect the initial results.

• Reads refine-logs/EXPERIMENT_PLAN.md (the best input), plus EXPERIMENT_TRACKER.md, FINAL_PROPOSAL.md or idea-stage files when present
• Implements experiment scripts, following the plan's run order: sanity, baseline, main method, then ablations
• Optionally runs a cross-model code review before deployment
• Deploys runs through the referenced run-experiment workflow and collects initial results for the auto review loop

Install the experiment-bridge skill

Download the skill ZIP from this page. You can install it in one of two ways, depending on how you use Claude.

For Claude Code, unzip the download into ~/.claude/skills/ to make it available in all projects, or into a project's .claude/skills/ folder to limit it to that project. After unzipping, the experiment-bridge folder should contain SKILL.md.

In the Claude apps, open Settings → Capabilities → Skills and upload the ZIP.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

What you need before you start

The skill assumes the earlier planning workflow already exists. Without an experiment plan there is nothing for it to read. It also deploys real compute jobs, so you need a GPU environment that the referenced run-experiment workflow can use.

• Claude Code or a compatible agent that reads the SKILL.md format
• Permission for Bash execution and reading and writing files
• A GPU environment for deployment, via the run-experiment workflow
• A written experiment plan, ideally refine-logs/EXPERIMENT_PLAN.md
• Optional: Codex, Antigravity, Grok or claude-review MCP servers for cross-model code review
• Optional: a wandb account, if you configure it in CLAUDE.md

How to trigger the skill with natural requests

You do not need a special command. The skill's description lists phrases that should activate it. You can say things like the examples below, in English or Chinese.

The skill also accepts an optional argument, the experiment plan path or a topic, as in the argument-hint in SKILL.md.

• "implement experiments"
• "deploy the plan"
• "bridge"
• "从计划到跑实验" or "实现实验"
• Or simply say you have an experiment plan ready to execute

Settings you can change

SKILL.md defines a few constants that control behaviour. You can override them in your request, for example by setting compact or base repo options.

• CODE_REVIEW = true: a second model reviews the code before deployment. Set false to skip.
• AUTO_DEPLOY = true: deploy automatically after implementation and review. Set false to inspect the code first.
• SANITY_FIRST = true: run the smallest sanity experiment before the rest.
• MAX_PARALLEL_RUNS = 4: maximum parallel deployments, limited by available GPUs.
• BASE_REPO = false: set a GitHub URL to implement experiments on top of an existing repo.
• COMPACT = false: when true, reads a compact idea summary and appends results to EXPERIMENT_LOG.md.

Safety notes: scripts, network and deployment

This skill runs arbitrary Bash commands, may clone repositories and may call external AI reviewer services. It also deploys compute jobs. Review its actions before letting it deploy, especially in production or sensitive environments. Setting AUTO_DEPLOY to false lets you inspect the code before anything runs.

Be aware that the skill is marked as running scripts and using the network. Check any repository it clones and any reviewer service it calls against your own policies.

• Set AUTO_DEPLOY = false to review code before deployment
• Keep CODE_REVIEW = true if you want a second model to check for logic bugs such as evaluation on another model's output
• Do not use it in production or sensitive environments without reviewing its actions

Tips for better results

The quality of the output depends on the quality of the plan. Make sure the plan lists datasets, compared systems, metrics, hyperparameters, seeds and a success criterion for each experiment. The skill checks its own code against these, so missing details become missing code.

Start with the sanity stage. It is the cheapest way to catch setup bugs before using GPU hours on the full run.

• Give each experiment a clear success criterion in the plan
• Keep SANITY_FIRST on to catch setup bugs early
• Check that FINAL_PROPOSAL.md describes the method you want implemented