Guide · Agent skill

How to use the ResearchClaw research pipeline agent skill

Updated 2026-10-29 · written with AI from the project's own documentation and checked against our catalogue data

The researchclaw skill lets an AI agent drive ResearchClaw, an autonomous research pipeline, from a single research topic. It is useful for researchers, students and developers who want a first-draft paper and experiment scaffold quickly. After this guide you will know what the skill does, how to install it, how to trigger it with a plain request, and which safety checks to make before running it.

ResearchClaw Autonomous Research Pipelineaiming-lab/AutoResearchClaw · MIT Download the skill

Quick start

1Install the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
2Install ResearchClaw. Make sure the ResearchClaw Python package and its researchclaw CLI are installed on your machine.
3Create config.yaml. Copy config.researchclaw.example.yaml to config.yaml if you do not have one.
4Add your LLM API key. Set llm.api_key in config.yaml or name an environment variable in llm.api_key_env.
5Choose an experiment mode. Set experiment.mode to simulated, sandbox or ssh_remote depending on where you want code to run.
6Make a request. Ask the agent to research a topic or run ResearchClaw by name, or run researchclaw run with --topic.
7Review the outputs. Open the output directory and check paper_draft.md, results_table.tex and the charts before relying on any content.

What is an agent skill?

An agent skill is a folder containing a SKILL.md file. The file tells an AI agent when to use a capability and how to perform it. Once a skill is installed, you can ask for the task in ordinary language and the agent follows the instructions in the file.

What the researchclaw skill makes the agent do

The skill runs ResearchClaw's 23-stage pipeline. Given a topic, the agent works through literature review, hypothesis generation, experiment design, code generation and execution, result analysis, paper writing, peer review and final export. The skill does not contain the pipeline itself. ResearchClaw is an external project, and the skill tells the agent how to operate it.

Outputs are saved to an output directory. The agent reports where items such as the paper draft, results tables and charts were written.

Experiment outputs: experiment.py, runs/run-1.json, experiment_summary.json
Paper outputs: paper_draft.md and results_table.tex
Charts: metric_trajectory.png and experiment_comparison.png
Overall status: pipeline_summary.json

Install the researchclaw skill

Download the ZIP from this page and unzip it. For a personal install, place the folder in ~/.claude/skills/. For a single project, place it in the project's .claude/skills/ folder.

In the Claude apps you can instead upload the skill under Settings → Capabilities → Skills.

Personal: ~/.claude/skills/researchclaw/
Project: .claude/skills/researchclaw/ inside your project
Claude apps: Settings → Capabilities → Skills, then upload

Install and prepare ResearchClaw

The skill requires the ResearchClaw Python package and its researchclaw CLI to be installed. The skill itself does not install them, so confirm they are available before you start. The agent checks for a config file and creates one from the example if needed.

Set your LLM API key either in config.yaml under llm.api_key or through an environment variable named in llm.api_key_env. Without a key the pipeline cannot connect to the model.

Requires the ResearchClaw Python package and researchclaw CLI
Requires an LLM API key in llm.api_key or llm.api_key_env
Requires Bash and file read/write access for the agent

ls config.yaml || ls config.researchclaw.example.yaml
cp config.researchclaw.example.yaml config.yaml

Choose an experiment mode

The experiment mode decides where generated code runs. Pick the most cautious mode that meets your needs.

simulated: the LLM generates synthetic results and no code is executed. Set experiment.mode: simulated.
sandbox: generated code runs locally via subprocess. Set experiment.mode: sandbox. Requires Python with numpy.
ssh_remote: generated code runs on a remote GPU server over SSH. Set experiment.mode: ssh_remote. Requires SSH access.

Trigger the skill with natural requests

The skill activates when you ask to research a topic, write a paper about a topic, investigate a topic, run an autonomous research pipeline, generate a research paper from scratch, or mention ResearchClaw by name. Include the topic and, if useful, the output location.

To run it directly from the command line, use the CLI. The --topic flag overrides the topic in config, --config points to a config file, and --output sets the output directory.

Example request: Research the effect of sparse attention on long-document summarisation and write a draft paper.
Example request: Use ResearchClaw to investigate low-cost soil moisture sensors.

researchclaw run --topic "Your research topic here" --auto-approve

Run the pipeline from Python

If you prefer code, the skill also documents a Python API. It loads the config, runs execute_pipeline with a run directory and run ID, and prints each stage's status. An iterative variant, execute_iterative_pipeline, runs multiple improvement rounds with max_iterations and convergence_rounds.

--from-stage resumes from a stage such as PAPER_OUTLINE
--auto-approve skips human approval at gate stages 5, 9 and 20

from researchclaw.pipeline.runner import execute_pipeline
from researchclaw.config import RCConfig
from researchclaw.adapters import AdapterBundle
from pathlib import Path

config = RCConfig.load("config.yaml", check_paths=False)
results = execute_pipeline(
    run_dir=Path("artifacts/my-run"),
    run_id="research-001",
    config=config,
    adapters=AdapterBundle(),
    auto_approve_gates=True,
)

for r in results:
    print(f"Stage {r.stage.name}: {r.status.value}")

Safety notes before you run it

The pipeline can execute generated code on your machine in sandbox mode, or on a remote server in ssh_remote mode. Review the config and the experiment mode first, and do not run it on a machine that holds sensitive data or credentials without that review. The skill also uses the network to reach your LLM provider.

The skill's own description warns that LLM-generated results and papers need expert verification before any use, because findings may be fabricated or unverified.

Run in simulated mode if you only need a draft structure
Check llm.base_url to confirm where your API key is sent
Keep credentials out of shared folders

Troubleshoot common problems

Most failures come from configuration. Validate the config first, then check the LLM connection and sandbox settings.

Config validation error: run researchclaw validate --config config.yaml
LLM connection failure: check llm.base_url and the API key
Sandbox failure: confirm experiment.sandbox.python_path exists and has numpy installed
Gate rejection: use --auto-approve or approve stages 5, 9 and 20 manually

researchclaw validate --config config.yaml

What people use it for

Drafting a first-pass paper outline and experiment scaffold on a new topic
Prototyping an automated research workflow as a developer with ResearchClaw already installed
Generating simulated experiment metrics and charts to test a figure layout
Running sandboxed experiment code to see what a generated script produces

Frequently asked questions

Is the researchclaw skill free to use?

The skill is published under the MIT licence. ResearchClaw itself is an external project, and the source does not state pricing for LLM API calls, which you pay to your provider.

What do I need to use the researchclaw skill?

You need the ResearchClaw Python package and researchclaw CLI, an LLM API key set in config.yaml or an environment variable, and Bash and file access for the agent. Sandbox mode also needs Python with numpy, and ssh_remote mode needs SSH access to a GPU server.

How do I install the researchclaw agent skill?

Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps, you can upload it under Settings → Capabilities → Skills.

Why does the pipeline fail with an LLM connection error?

The most common cause is an incorrect llm.base_url or API key. Check both in config.yaml, or confirm that the environment variable named in llm.api_key_env is set.

Can I trust the paper and results it generates?

No, not without expert checking. The skill notes that LLM-generated results and papers may be fabricated or unverified, so every claim, figure and metric should be verified before use.

How is this different from simply asking an AI chatbot to write a paper?

The skill runs a multi-stage pipeline that produces intermediate files, experiment code, results tables and charts in an output directory. A plain chat request usually returns only text. The pipeline is the external ResearchClaw project, which the skill tells the agent how to operate.

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

Use this skill if you are a researcher, student or developer who already has ResearchClaw and an LLM API key and wants a fast first draft and experiment scaffold. Pick something else if you need publication-ready research, or if you cannot review generated code and results before running them on a machine with sensitive data.

Download the skill