Guide · Agent skill
How to use the Full Research Pipeline agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Full Research Pipeline (skill name research-pipeline) is an agent skill that tells an AI agent to run a research project from start to finish: search for ideas in a research direction, plan and run experiments, get repeated automated reviews and fixes, then write up the results, optionally as a paper PDF. This guide explains what the skill does, how to install it, how to trigger it with plain requests, and what it needs and what to watch for before you let it run.
Full Research Pipelinewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download and unzip. Download the ZIP from this page and unzip it so the research-pipeline folder is ready to copy.
- 2Place the skill. Copy the folder into ~/.claude/skills/ for all projects, or into your project's .claude/skills/ for one project. Alternatively, upload it under Settings, Capabilities, Skills in the Claude apps.
- 3Install the referenced sub-skills. Make sure /idea-discovery, /experiment-bridge, /auto-review-loop, /paper-writing and the other referenced skills are present, since this file does not include them.
- 4Set up review access. Prepare Codex or GPT reviewer access for the default review route. The source does not describe the exact setup, so check your own provider's documentation.
- 5Start with a clear request. Ask for the full pipeline on a named research direction, and say whether you want it to stop before paper writing or set a venue.
- 6Watch the first checkpoints. Read the stage updates and NARRATIVE_REPORT.md as they appear, and interrupt the run if the direction or experiments look wrong.
What the skill makes the agent do
The skill is a set of orchestration rules in a SKILL.md file. It chains several stages in order: idea discovery, experiment bridging, an automated review loop, and optional paper writing. Each stage is handled by a referenced sub-workflow such as /idea-discovery, /experiment-bridge, /auto-review-loop and /paper-writing.
The pipeline keeps per-stage state under .aris/runs/ so that a run can resume from the first unfinished phase. It can also use a watchdog and iteration log to flag stalled phases or trigger a change of approach after repeated runs with no progress.
Check the requirements before you install
The skill needs Claude Code or another agent that can read SKILL.md files. It refers to sub-skills by path (/idea-discovery, /experiment-bridge, /auto-review-loop, /paper-writing, /research-lit, /render-html and /lean-formalize). These are not included in the file you download, so the pipeline will not work as a standalone file.
The default review route uses Codex or GPT reviewer access. The source implies external model access but does not say how to set it up. Optional extras include arXiv API access, a Feishu integration, and a chosen publication venue for paper writing.
- Sub-skills referenced by path must be present for the stages that use them
- Default review needs Codex or GPT reviewer access
- The skill runs scripts and uses the network
Install the skill
Download the ZIP for this skill from this page and unzip it. Put the research-pipeline folder into ~/.claude/skills/ to make it available to all your projects, or into a project's .claude/skills/ folder to limit it to that project.
You can also upload the skill in the Claude apps under Settings, then Capabilities, then Skills.
Trigger the skill with natural requests
The skill activates when your request matches its description. The source lists these trigger phrases: 全流程, full pipeline, 从找idea到投稿, end-to-end research, or a request for the complete autonomous research lifecycle. Name your research direction in the request so the agent knows the topic.
Because the pipeline runs for a long time, state clearly how far you want it to go: only ideas, through experiments, or all the way to a paper.
- Full pipeline on a topic: run the full pipeline on [your research direction]
- Stop at experiments: use the full pipeline on [topic] but stop before paper writing
- Paper target: end-to-end research on [topic], venue NeurIPS
Control the run with settings
The skill has constants that control behaviour. AUTO_PROCEED defaults to true, so checkpoints are progress updates and the run continues without waiting for you. AUTO_WRITE defaults to false, so the run stops after writing NARRATIVE_REPORT.md and you invoke /paper-writing yourself. VENUE has no default and is only needed when paper writing starts.
You can override these in the request. The source gives this example format:
- HUMAN_CHECKPOINT (default false) pauses each review round so you can give changes
- REVIEWER_DIFFICULTY (default medium) can be raised to hard or nightmare
- CODE_REVIEW (default true) reviews experiment code before it runs
- BASE_REPO (default false) builds experiments on a GitHub repo you name
/research-pipeline "topic" — AUTO_PROCEED: false, human checkpoint: true, difficulty: nightmare, code review: false, base repo: https://github.com/org/project, auto_write: true, venue: NeurIPSUnderstand what a run produces
Stage 1 searches the literature and generates ideas for your direction. Stage 2 plans and runs experiments, with optional code review before deployment. Stage 3 runs review and fix rounds. Stage 4 writes NARRATIVE_REPORT.md, and by default renders it to HTML with /render-html; a rendering failure does not stop the run.
Reviews in this pipeline are model-generated. The source notes that a same-family Codex review is recorded as provisional, and that results are only marked accepted after deterministic checks or a positive review from a known different model family.
- Stage 1: idea discovery and literature search
- Stage 2: experiment bridge
- Stage 3: automated review loop
- Stage 4: narrative report, optional HTML render
- Stage 5: optional paper writing once a venue is set
Safety notes before you run it
The skill can run scripts, use the network, launch experiments and set up schedulers on its own. The source says an overnight scheduler may check progress and nudge stalled phases, but it must never rerun or replace a reviewer verdict.
Do not point the pipeline at shared compute, credentials or live systems without someone supervising it. Review the experiment code (CODE_REVIEW is on by default) and keep the run interruptible, since you can stop a non-blocking run at any time.
- Keep credentials out of the project folder the agent can read
- Check which compute resources the experiments will use
- Read NARRATIVE_REPORT.md and the review record before trusting any result
What people use it for
- Scoping a new research direction by collecting relevant papers and candidate ideas before deciding what to test
- Running a set of experiments with review rounds that point out logic issues and suggest fixes
- Studying how a multi-stage autonomous agent pipeline is structured, using the SKILL.md orchestration as a worked example
- Producing a narrative report of a project's results to hand to a researcher for checking
Frequently asked questions
How do I install the research-pipeline skill?
Download the ZIP from this page and unzip it. Place the research-pipeline folder in ~/.claude/skills/ for all projects, or in a project's .claude/skills/ folder. You can also upload it in the Claude apps under Settings, Capabilities, Skills.
Is the Full Research Pipeline skill free?
The listing states the skill is released under the MIT licence. Any model access, such as Codex or GPT reviewer access, is separate and may cost money depending on your provider.
What do I need to run the full research pipeline?
You need Claude Code or a SKILL.md-compatible agent, the referenced sub-skills, and Codex or GPT reviewer access for the default review route. Optional extras are arXiv API access, Feishu integration, and a venue for paper writing.
Does the pipeline produce a peer-reviewed paper?
No. The reviews are generated by models, not peer reviewers, and results may be unverified. Treat the output as a draft to check yourself before any submission.
Why does the pipeline stop after the report and not write a paper?
By default AUTO_WRITE is false, so Stage 4 writes NARRATIVE_REPORT.md and stops. You can set auto_write to true and choose a venue to continue into paper writing.
How is this different from simply asking an agent to write a paper?
This skill runs several stages in order with state tracking, resumable runs and automated review loops. A single prompt usually lacks these checkpoints and records, though the source does not compare it to specific alternatives.
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 or developer who wants an agent to drive a long literature-to-experiments workflow with review checkpoints, and you can supply the referenced sub-skills and reviewer access. Pick something else if you need a quick answer, a verified or peer-reviewed paper, or an agent that will touch shared compute or live systems without supervision.
Download the skill