Skip to content
Get packs

Guide · Agent skill

How to use the Full Paper Writing Pipeline agent skill (paper-writing)

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

The paper-writing skill is a workflow file for Claude Code or a compatible SKILL.md agent. It tells the agent to take a research narrative or notes and run a chain of stages: outline planning, figure generation, LaTeX drafting, PDF compilation, and review-and-revise rounds. This guide explains what the skill does, how to install it, how to trigger it, what it needs, and where the risks are, so you can decide whether it fits your paper.

Full Paper Writing Pipelinewanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Confirm requirements. Check that Claude Code or a compatible SKILL.md agent is available, and that the sibling ARIS skills (paper-plan, paper-figure, paper-write, paper-compile, auto-paper-improvement-loop) and their shared references are installed.
  2. 2Install the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/, or upload it under Settings → Capabilities → Skills in the Claude apps.
  3. 3Prepare your narrative. Write a NARRATIVE_REPORT.md with claims, experiments, quantitative results and figure descriptions.
  4. 4Set the venue and options. Choose the target venue and illustration generator, and decide whether to pause between phases with AUTO_PROCEED or HUMAN_CHECKPOINT.
  5. 5Run the pipeline. Ask the agent in natural language or use /paper-writing with your narrative path, for example in the form shown in this guide.
  6. 6Check the output. Review the generated outline, figures, LaTeX and PDF, and verify every claim, citation and number against your own results.

What the skill makes the agent do

The skill is an orchestration layer. It chains five sub-skills in order: /paper-plan for the outline, /paper-figure for plots, /paper-write for LaTeX sections, /paper-compile to build the PDF, and /auto-paper-improvement-loop for review and polishing. Each phase builds on the previous phase's output, and the final deliverable is a paper/ directory with LaTeX source and a compiled PDF.

The visible skill file only defines this sequence. The detailed behaviour lives in the sibling skills, which are not included in this file. The skill can also call external AI reviewer models through MCP tools.

/paper-plan → /paper-figure → /paper-write → /paper-compile → /auto-paper-improvement-loop

Install the paper-writing skill

Option 1: download the ZIP from this page and unzip it into ~/.claude/skills/ for a personal install, or into a project's .claude/skills/ folder to scope it to one project. The skill's file is SKILL.md.

Option 2: in the Claude apps, open Settings → Capabilities → Skills and upload the skill there.

Installing this file alone is not enough to run the full pipeline. You also need the sibling ARIS skills listed under requirements below.

  • Personal install: ~/.claude/skills/
  • Project install: .claude/skills/ inside your project
  • Claude apps: Settings → Capabilities → Skills

Trigger the skill with natural requests

The skill's description lists the phrases that should activate it. You can ask in English or Chinese. Describe your input and the target, and the agent should pick up the workflow.

You can also pass inline options. The source shows the form /paper-writing with a narrative file path followed by settings such as venue and illustration mode.

  • "paper writing"
  • "write paper pipeline"
  • "写论文全流程"
  • "从报告到PDF"
  • Ask for the complete paper generation workflow from a narrative report
/paper-writing "NARRATIVE_REPORT.md" — venue: NeurIPS, illustration: gemini, human checkpoint: true

Choose inputs and key settings

The pipeline accepts a NARRATIVE_REPORT.md (the best input: claims, experiments, results and figures), research direction plus experiment results (the agent drafts the narrative first), or an existing PAPER_PLAN.md, which skips the planning phase. The more detail you give, especially figure descriptions and quantitative results, the better the output.

Main settings in the source include the target venue (for example ICLR, NeurIPS, ICML, ACL, IEEE_JOURNAL or IEEE_CONF), the number of improvement rounds (default 2), the reviewer model (default gpt-6-astra via Codex MCP), and the illustration generator. The default illustration option is figurespec, which produces deterministic diagrams. Options are gemini, codex-image2, mermaid, or false to skip the step.

  • Default venue: ICLR
  • Default improvement rounds: 2
  • Default illustration generator: figurespec
  • HUMAN_CHECKPOINT = true pauses after each review round so you can give instructions

Optional: match the structure of a reference paper

Add — style-ref: <source> to steer structure such as section order, theorem density and bibliography style toward a paper you admire. It is off by default. The source accepts a local TeX directory or file, a local PDF, an arXiv id, or an http(s) URL. Overleaf URLs and IDs are rejected; clone the project locally first and pass that path.

The style profile is meant as structural guidance only. The source says the agent must not copy prose, claims, examples or terminology from the reference, and must not pass the style reference to the reviewer steps.

Safety notes: scripts and network

This skill runs scripts and uses the network. Its allowed tools include Bash, file read and write tools, and MCP tools for Codex, Antigravity, Grok and Claude review. The review steps send paper content to external reviewer models. If your results are confidential or unpublished, check the data policy of each reviewer you route to before running the pipeline.

The file gives no verification step for facts. Claims, citations and numbers in the output must be checked by you before submission.

  • Review phases send paper content to external reviewer models
  • Check the data policy of any third-party reviewer before using unpublished results
  • Verify every claim, citation and number yourself

Tips for a better result

Write a detailed narrative first. Give exact quantitative results and clear figure descriptions, because the output can only be as specific as the input. Run with AUTO_PROCEED set to false if you want to approve each phase. Turn on HUMAN_CHECKPOINT if you want to steer the improvement rounds yourself.

Keep the sibling skills and reviewer MCP server set up before you start, so the pipeline does not stop partway through.

  • Start with a NARRATIVE_REPORT.md that has claims, results and figure notes
  • Use AUTO_PROCEED: false to approve phases one at a time
  • Use HUMAN_CHECKPOINT: true to review scores and steer each round

What people use it for

  • Turning a finished set of experiment results and a narrative report into a first complete draft with LaTeX source and PDF.
  • Producing a conference-formatted draft for a venue such as ICLR or NeurIPS by setting the VENUE option.
  • Students who want structured help organising a submission and running revision rounds.
  • Teams already using the ARIS skill set and a Codex reviewer, who want one command to run the full chain.

Frequently asked questions

Is the paper-writing skill free?

The skill file is published under the MIT licence. The source does not give prices for the sibling skills or reviewer services, so check those separately. Reviewer models called through MCP may have their own costs and terms.

What do I need to install before using it?

You need Claude Code or a compatible SKILL.md agent, the sibling ARIS skills (paper-plan, paper-figure, paper-write, paper-compile and auto-paper-improvement-loop) with their shared references, and a LaTeX toolchain for compilation. You also need a reviewer MCP server such as Codex, plus Python 3.

Why does the skill fail when I run it alone?

The pipeline calls /paper-plan, /paper-write and other skills that are not included in this file. Without them installed, the pipeline cannot run standalone. Install the sibling skills first.

Do I need a GEMINI_API_KEY?

Only if you choose the gemini illustration option. The default illustration option is figurespec, and the source does not require a Gemini key for it.

Will my unpublished results be sent to third parties?

The review steps send paper content to external reviewer models. Check the data policy of the reviewer you configure before using confidential or unpublished results.

Can I use a reference paper to guide the style?

Yes, with the optional — style-ref: <source> argument. It only guides structure, such as section order and citation style, and it is kept away from the reviewer steps. It is off unless you pass it.

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 already have experiment results and a narrative, and you want an end-to-end draft with LaTeX, a PDF and automated review rounds, and you can install the sibling ARIS skills. Pick something else if you want a finished, submission-ready paper without author review, if you cannot install the sibling skills, or if you cannot send unpublished results to external reviewer models.

Download the skill

Find exactly what
you're looking for