How to use the Paper Writing Pipeline agent skill

What an agent skill is and what this one does

An agent skill is a folder containing a SKILL.md file that tells an AI coding agent how to carry out a specific workflow. When the skill matches your request, the agent follows its instructions instead of improvising.

This skill, named paper-writing, is labelled 'Workflow 3: Full paper writing pipeline'. It chains five sub-skills: paper-plan, paper-figure, paper-write, paper-compile and auto-paper-improvement-loop. Together they take a narrative report to a polished PDF with LaTeX source in a paper/ directory.

• Phase 1 (paper-plan): parses your report for claims, evidence and figures, and builds a Claims-Evidence Matrix and section outline in PAPER_PLAN.md.
• Phase 2 (paper-figure): generates matplotlib or seaborn plots and LaTeX tables from your JSON or CSV data, and creates figures/latex_includes.tex.
• Phase 3 (paper-write): drafts section-by-section LaTeX with citations, builds references.bib, and runs an automated bib cleaning and de-AI wording pass.
• Phase 4 (paper-compile): runs latexmk -pdf, tries up to 3 times to fix common errors, and checks undefined references, page count and font embedding.
• Phase 5 (auto-paper-improvement-loop): runs two review-fix-recompile rounds on the full paper.

Requirements before you install

The skill depends on several other pieces. Check these first, because the pipeline will not run cleanly without them.

The five sub-skills are not contained in this file. They must be installed separately, after the skills/skills-codex/* package. The reviewer step needs a local gemini-review MCP bridge. The compile step needs Python plotting libraries (matplotlib and seaborn) and a LaTeX toolchain.

• Dependent skills: paper-plan, paper-figure, paper-write, paper-compile and auto-paper-improvement-loop.
• Local gemini-review MCP bridge, which hands the review steps to Gemini.
• Python with matplotlib and seaborn for plots.
• A LaTeX toolchain that supports latexmk and pdftotext.
• GEMINI_API_KEY environment variable, only if you turn on the optional illustration step (illustration: true).

Install the paper-writing skill

Download the ZIP from our page and unzip it. You can put the skill in one of two places, depending on whether you want it everywhere or only in one project.

For a personal install that applies to all your projects, unzip it into ~/.claude/skills/. For a project-only install, unzip it into the .claude/skills/ folder inside that project.

If you use the Claude apps instead of a local agent, open Settings, go to Capabilities, then Skills, and upload the package there. Remember to install the dependent skills too, after the skills/skills-codex/* package.

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

Trigger the skill with natural requests

You do not need a special command. The skill's description lists phrases that should activate it. Say one of these, along with your report or a description of your work:

The skill also accepts a direct slash-style call with the report file name as the argument. The source shows an inline override that sets the venue and the human checkpoint in the same line.

• "write paper pipeline"
• "paper writing"
• "从报告到PDF" (from report to PDF)
• "写论文全流程" (full paper writing workflow)
• Or ask for the complete paper generation workflow in your own words.

What you need to provide as input

The pipeline accepts one of three inputs. The best is a NARRATIVE_REPORT.md file with structured claims, experiments, results and figure descriptions. If you do not have one, give it a research direction and your experiment results, and the skill will help draft the narrative first. If you already have a PAPER_PLAN.md, the skill skips Phase 1 and starts at Phase 2.

The source states that the more detailed the input is, especially figure descriptions and quantitative results, the better the output.

• Best: NARRATIVE_REPORT.md with claims, experiments, results and figures.
• Alternative: research direction plus experiment results.
• Shortcut: an existing PAPER_PLAN.md, which skips Phase 1.

Settings you can change

The skill has a small set of constants. You can override them inline when you call it, as shown above.

VENUE defaults to ICLR and can also be NeurIPS or ICML, which affects the style file, page limit and citation format. MAX_IMPROVEMENT_ROUNDS defaults to 2. AUTO_PROCEED defaults to true, so phases continue without waiting; set it to false to pause for your approval after each phase. HUMAN_CHECKPOINT defaults to false; set it to true to see the review score after each improvement round and give custom instructions.

• VENUE: ICLR (default), NeurIPS or ICML
• MAX_IMPROVEMENT_ROUNDS: 2 by default
• AUTO_PROCEED: true by default; false pauses after each phase
• HUMAN_CHECKPOINT: false by default; true pauses each improvement round
• illustration: optional AI illustration step in Phase 2b, off unless set

Safety notes: network use and scripts

This skill runs scripts and uses the network. The reviewer step sends your paper content to an external model service through the gemini-review bridge. If your data or results are unpublished or confidential, do not run this skill on them.

The skill also writes files into a paper/ directory and a figures/ directory, runs latexmk and pdftotext, and generates plotting scripts. Read the output before you rely on it.

• Reviewer step: content goes to an external model service via the gemini-review MCP bridge.
• Do not use it on unpublished data or confidential results that must stay on your machine.
• The illustration step needs GEMINI_API_KEY and is off by default.