Guide · Agent skill
How to use the research-pipeline agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
research-pipeline is an agent skill that tells an AI coding agent to run a research project in stages, from a broad research direction through idea discovery and experiments to review and a narrative report that can feed into paper writing. This guide explains what the skill makes the agent do, how to install it, how to trigger it with a plain request, and what it needs before you run it. By the end you will know how to start a run, how to resume an interrupted one, and which settings control how much runs without you.
Full Research Pipelinewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Confirm the requirements. Check that you have Claude Code or a compatible agent, the Codex MCP server, web search and fetch access, shell access, and the other skills from the same repository. Without the reviewer and compute setup, the run can stall partway.
- 2Install the skill folder. Download the ZIP from the page and unzip it into ~/.claude/skills/ for personal use or into the project's .claude/skills/ folder for one project. In the Claude apps, upload it under Settings → Capabilities → Skills.
- 3Choose your settings. Decide whether you want checkpoints to pause, whether to review each round, and whether paper writing should start automatically. Leave the defaults if you want a long run with few stops.
- 4Start with a natural request. Ask for the full pipeline on a research direction using a phrase such as "full pipeline" or "end-to-end research". Add a topic that is specific enough to survey.
- 5Review the outputs. When the run finishes Stage 4, read NARRATIVE_REPORT.md and the review results. Check the claims and experiments yourself before you use them.
- 6Resume if interrupted. If the run stops, use the resume flag with its run ID. The skill re-checks any stage that finished writing but was not yet accepted.
What the research-pipeline skill makes the agent do
An agent skill is a folder of instructions that an AI coding agent such as Claude Code loads when a request matches it. This skill covers the full research lifecycle in five stages: literature survey from a broad research direction, idea generation and filtering, experiments, review of the work by a second model, and a narrative report.
Progress is saved to a run-state file, so an interrupted run can be resumed instead of restarted. Several stages are delegated to other skills in the same repository, which the skill references but does not include in the file you download.
- Stage 1 surveys literature for the research direction.
- Stage 2 generates and filters ideas.
- Stage 3 runs experiments and the review loop.
- Stage 4 produces NARRATIVE_REPORT.md, which can feed into paper writing.
- Stage 5 is paper writing, which starts only when a target venue is set.
Install the research-pipeline skill
You can install the skill in one of two ways. For a personal setup, download the ZIP from the page, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder instead.
If you use the Claude apps rather than Claude Code, upload the skill under Settings → Capabilities → Skills.
The download contains SKILL.md. The skill also depends on other skills from the same repository, so a bare install will not cover every stage.
- Personal install: ~/.claude/skills/
- Project install: the project's .claude/skills/ folder
- Claude apps: Settings → Capabilities → Skills
~/.claude/skills/How to trigger the skill with a natural request
You do not need a command to start the skill. The skill loads when your request matches its description, which includes phrases such as "full pipeline", "end-to-end research", "全流程", "从找idea到投稿", or a request for the complete autonomous research lifecycle.
If you prefer an explicit call, the skill accepts a research direction as its argument. To resume an earlier run, pass its run ID with the resume flag. Settings such as human checkpoints and the reviewer difficulty can be passed in the same request.
- "Run the full pipeline on sparse attention for long-context models."
- "Start an end-to-end research run on graph-based anomaly detection."
- "Resume research run 2024-run-03 from where it stopped."
/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: NeurIPSSettings that control how much runs unattended
The skill has a small set of constants at the top of SKILL.md. The defaults favour a long run with few stops, so review them before you start.
AUTO_PROCEED defaults to true, which means checkpoints report their result and the run continues in the same turn. Set it to false if you want the agent to stop and ask at each selection point.
HUMAN_CHECKPOINT defaults to false. Set it to true to pause after each review round so you can see the score and give modification instructions. RESUMABLE defaults to true, which records per-stage state so you can resume.
AUTO_WRITE defaults to false, so the run stops after Stage 4 and you start paper writing yourself. VENUE has no default and is only needed when Stage 5 begins; it is never guessed.
- AUTO_PROCEED: true by default; checkpoints are informational.
- HUMAN_CHECKPOINT: false by default; enable it to review each round.
- AUTO_WRITE: false by default; paper writing is started manually.
- VENUE: unset; choose from the listed venues when Stage 5 begins.
- REVIEWER_DIFFICULTY: medium by default; hard and nightmare add reviewer memory and debate.
Safety notes: scripts, network and sensitive data
The skill allows shell commands, file writes, web search and web fetch. It runs experiment code, so it can execute scripts on your machine. Its second-model reviewer runs through the Codex MCP server, so content from the run is sent to that external service.
Automated results can be wrong or overstated. The skill itself states that quality verdicts rest on its reviewer loop, so check results yourself before you rely on them. Do not use the skill with sensitive or proprietary data unless you are comfortable with web search, web fetch and the external reviewer seeing it.
Heartbeat or overnight scheduling can nudge a stalled phase forward, but it cannot decide that the work is good enough. That judgement stays with the skill's review loop and with you.
- Runs scripts: yes, including experiment code.
- Uses network: yes, for web search, web fetch and the reviewer.
- Review outputs yourself before using any claim.
- Keep proprietary or sensitive data out of runs unless the external services are acceptable.
What people use it for
- Surveying literature on a new research direction and shortlisting ideas for experiments.
- Running an exploratory experiment cycle on a topic overnight, then reviewing the results the next day.
- Testing whether an automated pipeline can shorten the time from idea to experiment in an ML or science team.
- Producing a narrative report that a researcher can turn into a draft with the paper-writing skill.
Frequently asked questions
Is the research-pipeline skill free?
The skill is published under the MIT licence, as stated on the page. Running it still requires Claude Code or a compatible agent, the Codex MCP server, and any compute you use for experiments, which may have their own costs.
What do I need to install research-pipeline?
Download the ZIP and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it in the Claude apps under Settings → Capabilities → Skills. The skill also needs the other skills from its repository, the Codex MCP server, web access and shell access.
How do I resume an interrupted research-pipeline run?
Pass the resume flag with the run ID, for example /research-pipeline — resume <run_id>. The skill records per-stage state in a run-state file under .aris/runs/ and re-validates any stage that finished but was not accepted.
Does research-pipeline need Codex MCP?
Yes, for the default cross-model reviewer, which uses the Codex MCP server (mcp__codex__codex). The skill also has settings that route review differently, but the default path depends on Codex MCP.
Can research-pipeline write a publishable paper by itself?
Not reliably. The skill's own description says quality verdicts rest on its reviewer loop, and results still need human checking because automated claims can be wrong or overstated. Paper writing starts only when AUTO_WRITE is enabled or when you run the paper-writing skill yourself.
What is an alternative to research-pipeline?
If you only need one part of the lifecycle, use the individual skills in the same repository, such as research-lit for literature or auto-review-loop for review. If you need a fully manual process, run each stage yourself and use the pipeline only for drafting.
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 research-pipeline if you want a long, mostly unattended exploratory run on a topic, have Codex MCP and compute set up, and will review every output afterwards. Pick something else if you need a publication-ready paper without expert review, lack the reviewer or compute setup, or must keep sensitive data away from external web and reviewer services.
Download the skill