Guide · Agent skill
How to use the Research Refine Pipeline agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The research-refine-pipeline skill is an agent skill for Claude Code or a compatible agent that runs a full research-planning workflow in one pass. It takes a rough research direction, refines it into a focused method proposal, reviews it, and then writes a detailed experiment plan, a tracker and a short summary of what to run next. This guide explains what the skill does, how to install it, how to trigger it, and what it needs before you start.
Research Refine Pipelinewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download the ZIP. Download the skill ZIP from this page. It contains SKILL.md.
- 2Unzip into a skills folder. Unzip into ~/.claude/skills/ for personal use, or into .claude/skills/ inside a project. In the Claude apps, upload it under Settings → Capabilities → Skills instead.
- 3Check the dependencies. Confirm the sibling research-refine and experiment-plan skills and the shared reference files are available, and that a Codex MCP server or another reviewer backend is configured.
- 4Start a new session. Open a new agent session in your project so the skill is loaded.
- 5Describe your research direction. Make a request with the problem, rough approach and constraints, for example asking for an end-to-end pipeline.
- 6Review the refine-logs folder. Read FINAL_PROPOSAL.md, EXPERIMENT_PLAN.md and PIPELINE_SUMMARY.md before acting on them.
- 7Run experiments separately. The summary points to /run-experiment as the next step. The pipeline itself only plans.
What is an agent skill, and what does this one do?
An agent skill is a folder containing a SKILL.md file that tells an AI agent how to carry out a specific kind of task. When a request matches the skill's description, the agent follows its instructions. This skill is named research-refine-pipeline and is built to chain two sibling skills, research-refine and experiment-plan, so you do not have to run them one after another.
The workflow starts from a vague research idea and produces a package: a problem-anchored final proposal, a review history explaining why the method is focused, a claim-driven experiment roadmap, and a compact pipeline summary. It plans experiments but does not run them.
- Refines the method first, then plans experiments, so the plan is built on a stable thesis.
- Writes its review history and claims to files, so you can check the reasoning later.
- Can call reviewer models through a configurable reviewer route; by default this is Codex via MCP.
Install the research-refine-pipeline skill
You can install the skill in one of three places. For a personal setup, download the ZIP from the skill's page on this directory and unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.
The download contains one file, SKILL.md. After unzipping, the skill folder should contain that file. Start a new agent session so the skill is picked up.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project
- Claude apps: Settings → Capabilities → Skills, then upload
What the skill needs before you run it
The skill needs more than the SKILL.md file. It depends on two sibling skills, research-refine and experiment-plan, and on shared reference files that the instructions reference by relative paths. The files for those shared references are not included in this download, so the skill may not work fully without them.
The agent also needs tools: Bash, file read and write, search, and web tools. For reviewer passes, it expects a Codex MCP server (mcp__codex__codex) or another configured reviewer backend. The skill runs scripts and uses the network.
- Claude Code or a compatible agent with Bash, file read/write, search and web tools
- Codex MCP server (mcp__codex__codex) or another configured reviewer backend
- Sibling skills research-refine and experiment-plan
- Shared reference files referenced by relative paths (not included in this download)
How to trigger the skill with natural requests
The skill activates when your request matches its description. The description mentions going from a vague research direction to a final proposal plus experiment roadmap, building a pipeline, doing something end-to-end, and generating the method and experiment plan together. The skill also responds to the phrase 串起来, which means chaining things together.
Be specific about your starting point. Say what the problem is, what rough approach you have in mind, and any constraints or target venue. The skill's first phase extracts these details.
- "Take this rough idea on sample-efficient retrieval and give me a final proposal plus a full experiment roadmap."
- "Do the whole thing end-to-end: refine my method and plan the experiments."
- "串起来: refine this method and generate both the method and experiment plan together."
What output files to expect
The skill writes its outputs into a refine-logs/ folder. The main deliverables are FINAL_PROPOSAL.md, REVIEW_SUMMARY.md, REFINEMENT_REPORT.md, EXPERIMENT_PLAN.md, EXPERIMENT_TRACKER.md and PIPELINE_SUMMARY.md. The summary file lists the dominant contribution, the must-prove claims, the first runs to launch, the main risks and a next action.
When the work is done, the agent reports the locations of these files and suggests /run-experiment as the next step. Running experiments is a separate step.
- refine-logs/FINAL_PROPOSAL.md
- refine-logs/REVIEW_SUMMARY.md
- refine-logs/REFINEMENT_REPORT.md
- refine-logs/EXPERIMENT_PLAN.md
- refine-logs/EXPERIMENT_TRACKER.md
- refine-logs/PIPELINE_SUMMARY.md
Safety notes: scripts and network access
This skill runs scripts and uses the network. Its allowed tools include Bash, WebSearch, WebFetch and the Codex MCP tools. Before you run it on a project with sensitive material, check what it will read and what it may send to a reviewer backend, and confirm which reviewer route is configured.
The skill also writes files and may rewrite them in chunks if a write fails due to file size. Review the refine-logs/ folder after each run.
- Bash and web tools are allowed, so review commands the agent proposes if your setup asks for approval.
- Reviewer passes may send your proposal text to the configured reviewer backend.
- Check the output files before relying on them for a paper.
Tips for better results
Start with a clear problem statement, since the first phase checks whether a stored proposal still matches your request. If a proposal already exists and is outdated, say so, so the agent re-runs the refinement stage instead of planning experiments for the wrong method.
If your method is intentionally simple, the skill is designed to defend that simplicity rather than add components. If you only need one stage, use research-refine or experiment-plan directly.
- State the problem, rough approach, constraints and target venue up front.
- If an old FINAL_PROPOSAL.md exists, say whether it still applies.
- Use the staged skills when you only need method refinement or only experiment planning.
What people use it for
- Turning a rough research idea into a focused method proposal with review history.
- Drafting a paper plan that includes a claim-driven experiment roadmap and run order.
- Producing a tracker and first-runs list to hand to a team after planning.
- Chaining the research-refine and experiment-plan skills when a team already uses both.
Frequently asked questions
How do I install the research-refine-pipeline 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 instead.
What does the research-refine-pipeline skill require?
It needs Claude Code or a compatible agent with Bash, file, search and web tools, plus a Codex MCP server or another reviewer backend. It also depends on the sibling research-refine and experiment-plan skills and on shared reference files that are not included in this download.
Is the research-refine-pipeline skill free?
The skill is listed under the MIT licence. Check the licence file in the source repository for the terms that apply to your use.
Does the skill run experiments for me?
No. It plans the method and experiments and writes a summary. The summary's next action points to /run-experiment, which is a separate step.
What is the difference between this skill and research-refine or experiment-plan?
research-refine handles method refinement only, and experiment-plan handles experiment planning only. This pipeline chains both into one run and also writes a pipeline summary with a run order.
Why does my skill not work after installing it?
Check that the sibling research-refine and experiment-plan skills are installed, that the shared reference files referenced by relative paths are present, and that a reviewer backend such as Codex MCP is configured. Then start a new session so the skill is loaded.
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 want a complete method proposal and experiment roadmap from a rough idea, and you already have, or are willing to set up, the sibling skills and a reviewer backend. Pick something lighter for simple one-off questions, or if you need finished experiments, since this skill only plans and does not run them.
Download the skill