Guide · Agent skill
How to use the Experiment Bridge agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Experiment Bridge is an agent skill for machine-learning research workflows. It reads a written experiment plan (EXPERIMENT_PLAN.md), implements the experiment code, deploys the runs to a GPU and collects first results. This guide explains what an agent skill is, how to install this one, how to trigger it, and what to check before letting it spend compute. By the end you will know how to prepare the plan file, start the skill and review what it produced.
Experiment Bridgewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download the ZIP. Get the experiment-bridge ZIP from the skill's page on this directory.
- 2Unzip the skill. Extract it into ~/.claude/skills/ for personal use, or into your project's .claude/skills/ folder.
- 3Prepare the plan. Make sure refine-logs/EXPERIMENT_PLAN.md exists and lists experiments, metrics, hyperparameters and seeds.
- 4Configure GPU and reviewer settings. Set your compute backend and the reviewer configuration the file references, and set AUTO_DEPLOY to false if you want to review code first.
- 5Ask for implementation. Send a request such as "Implement the experiments in the plan and deploy them" to the agent.
- 6Review the code and results. Check the generated scripts against the plan, then read the JSON or CSV outputs before using them.
What is an agent skill, and what does this one do?
An agent skill is a folder containing a SKILL.md file with instructions that a compatible AI agent (Claude or Codex) loads when your request matches the skill's description. The skill tells the agent how to carry out a specific workflow.
Experiment Bridge sits between idea generation and an automated review loop. It takes the output of an earlier planning step and turns it into running experiments.
- Reads refine-logs/EXPERIMENT_PLAN.md and related files such as EXPERIMENT_TRACKER.md and FINAL_PROPOSAL.md
- Writes experiment code: training scripts with argparse, evaluation scripts, data loading, baselines, fixed random seeds, and results saved as JSON or CSV
- Optionally asks a second AI reviewer to check the code for logic bugs before deployment
- Deploys runs to GPU compute and collects initial results for the next workflow stage
Install the experiment-bridge skill
Download the ZIP from the skill's page on this directory, then choose one of the two install routes below. The downloaded package contains the SKILL.md file for this skill.
For a personal install that applies to all your projects, unzip the file into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder. In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into your project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
Prepare the experiment plan before you trigger it
The skill works best when an experiment plan already exists. It expects refine-logs/EXPERIMENT_PLAN.md, typically produced by an earlier experiment-planning workflow. The source also lists EXPERIMENT_TRACKER.md, FINAL_PROPOSAL.md and idea-stage files as optional inputs.
If no plan file exists, the skill asks you what experiments to implement. The plan should specify datasets, compared systems, metrics, hyperparameters, seeds and a success criterion for each experiment.
- Keep the plan in refine-logs/EXPERIMENT_PLAN.md inside your project
- Mark each experiment as MUST-RUN or NICE-TO-HAVE
- Give the compute budget in estimated GPU-hours
- Specify the GPU backend you want (local, SSH server, Vast.ai or Modal) and any wandb setup in AGENTS.md
Trigger the skill with a natural request
You do not need a special command. The skill loads when your request matches its description. The source lists these trigger phrases: "实现实验", "implement experiments", "bridge", "从计划到跑实验", "deploy the plan", or any request where you have an experiment plan ready to execute.
The source also shows an override syntax that sets constants for one run. Its example reads: /experiment-bridge "EXPERIMENT_PLAN.md" — compact: true, base repo: https://github.com/org/project
- "Implement the experiments in refine-logs/EXPERIMENT_PLAN.md"
- "Deploy the plan and collect initial results"
- "Bridge the experiment plan to running experiments"
- "Implement experiments from the plan, but don't deploy until I review the code"
/experiment-bridge "EXPERIMENT_PLAN.md" — compact: true, base repo: https://github.com/org/projectConfigure the key settings
The skill has constants that control its behaviour. The most important is AUTO_DEPLOY. When it is true, the agent deploys experiments automatically after implementing them. Set it to false if you want to read the code before anything runs.
CODE_REVIEW runs a secondary reviewer on the experiment code before deployment. SANITY_FIRST runs the smallest sanity-stage experiment first to catch setup bugs early. MAX_PARALLEL_RUNS limits parallel deployments, which is capped by how many GPUs you have.
- AUTO_DEPLOY = true: deploy automatically after implementation; false to review first
- CODE_REVIEW = true: a secondary reviewer checks code for logic bugs before deployment
- SANITY_FIRST = true: run the sanity experiment first
- MAX_PARALLEL_RUNS = 4: maximum experiments in parallel
- BASE_REPO = false: set a GitHub URL to implement on top of an existing codebase
- COMPACT = false: when true, reads the compact idea summary and appends runs to EXPERIMENT_LOG.md
Check the results and safety notes before scaling up
Read the generated code before you trust its numbers. The skill's self-review checks that evaluation compares predictions against dataset ground truth, not another model's output. Confirm this yourself, along with seeds, hyperparameters and result file paths.
Running with AUTO_DEPLOY enabled means the agent deploys code automatically. The source warns that this can spend GPU hours and money on buggy code. The skill runs scripts and uses the network, so only use it on machines and accounts you control.
- Check that every hyperparameter from the plan appears in argparse
- Confirm the random seed is fixed and settable
- Verify results are written to JSON or CSV
- Use AUTO_DEPLOY = false on shared or paid GPU infrastructure until the code is reviewed
- If you use wandb, the skill only logs when it is configured in AGENTS.md
What people use it for
- Turning a written EXPERIMENT_PLAN.md with baselines and ablations into training and evaluation scripts
- Scaffolding a machine-learning experiment with argparse flags, fixed seeds and JSON or CSV outputs
- Running sanity-stage experiments first to catch setup bugs before launching the full grid
- Preparing first results for a later automated review loop
Frequently asked questions
Is the Experiment Bridge skill free?
The source lists the licence as MIT. It does not state pricing for the skill itself. GPU compute on local, SSH, Vast.ai or Modal is set up by you and may have its own cost.
How do I install the experiment-bridge skill?
Download the ZIP from the skill's page, then unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps you can upload the ZIP under Settings → Capabilities → Skills.
What do I need to run it?
You need a Codex-compatible or Claude-compatible agent environment, the companion skills and reviewer configuration the file references, and GPU compute. A wandb account is optional and only needed if you configure it in AGENTS.md.
Can I use it without an experiment plan?
The skill expects an EXPERIMENT_PLAN.md from an earlier workflow. Without one, it asks what experiments to implement, so you should write or generate a plan first.
Why did the agent deploy buggy code?
With AUTO_DEPLOY enabled, the agent deploys code automatically. Set AUTO_DEPLOY to false to review the code before it runs, and keep CODE_REVIEW enabled to have a second reviewer check for logic bugs.
How is this different from run-experiment?
The source describes Experiment Bridge as a bridge that implements code and deploys it, and it references /run-experiment for deployment. The source does not describe run-experiment's internals, so check that skill's own documentation for details.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Experiment Bridge suits researchers and students who already have a written experiment plan and want an agent to write and launch machine-learning experiments on GPUs. It is a poor fit for projects without a plan file, or for anyone who wants code deployed to shared or paid GPUs without reviewing it first.
Download the skill