How to use the Paper Write (LaTeX Drafting) agent skill

What is an agent skill and what does paper-write do?

An agent skill is a folder containing a SKILL.md file that tells an AI coding agent how to carry out a specific task. When the skill is installed and the agent judges it relevant to your request, it follows the instructions in that file.

The paper-write skill draws a LaTeX paper from a paper plan. It expects a PAPER_PLAN.md with a claims-evidence matrix, section plan and figure plan, plus a NARRATIVE_REPORT.md that holds the research narrative. From these it creates a paper/ directory with main.tex, a venue style file, math_commands.tex, references.bib, a sections/ folder and a figures/ folder.

• Writes each section in order, drawing content from the narrative and the plan.
• Inserts figures and tables using snippets from figures/latex_includes.tex when that file exists.
• Uses \citep and \citet for ML venues, and \cite for IEEE venues.
• Fetches real BibTeX entries from DBLP/CrossRef by default instead of generating citations from memory.
• Backs up an existing paper/ folder before overwriting it and removes section files that are no longer referenced.

Install the paper-write skill

Download the ZIP for this skill from this page and unzip it. To make the skill available to Claude Code for all your projects, place the folder in ~/.claude/skills/. To limit it to one project, place it in that project's .claude/skills/ folder.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

If you plan to use the Codex-reviewer overlay, install the skills-codex package first, then install this package after it, as the source notes.

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

Prepare your inputs before you ask

The skill works from files, so the quality of your draft depends on what you give it. Have PAPER_PLAN.md and NARRATIVE_REPORT.md ready in your project. These are normally produced by companion skills such as /paper-plan. If you have generated figures, the /paper-figure skill produces figures/latex_includes.tex, which the writing step references.

If PAPER_PLAN.md is missing, the skill will ask you to run /paper-plan first or to provide a brief outline. You will also need LaTeX style files or templates, which the skill looks for in its templates/ folder. For the default bibliography workflow, the agent needs network access to DBLP and CrossRef.

• PAPER_PLAN.md (required)
• NARRATIVE_REPORT.md (required)
• figures/ with PDF or PNG outputs and figures/latex_includes.tex (optional but referenced)
• An existing .bib file, or the skill will create one
• Network access to DBLP/CrossRef for the default citation workflow

Trigger the skill with a natural request

The skill description lists phrases that should activate it, including "write paper", "draft LaTeX", "开始写", "写论文", or any request to generate LaTeX content from a paper plan. You can also invoke it directly with the /paper-write command, which takes your paper topic as its argument.

Be specific about the venue and the paper's focus so the agent picks the right template and section structure.

• "Write paper on the sparse routing method using PAPER_PLAN.md, target NeurIPS."
• "Draft LaTeX from my plan for an IEEE conference submission, no anonymous author block."
• "Start writing the paper from the outline in this folder."

Choose a venue and key settings

The skill has defaults you can change in your request. TARGET_VENUE defaults to ICLR. Supported values include ICLR, NeurIPS, ICML, CVPR, ACL, AAAI, ACM, IEEE_JOURNAL and IEEE_CONF. ANONYMOUS defaults to true, so the author block reads Anonymous Authors; set it to false for camera-ready versions. Note that most IEEE venues do not use anonymous submission.

Page limits also differ. For ML conferences, references and appendix are not counted toward the main body limit. For IEEE venues, references count toward the limit.

• ML conferences use natbib (\citep, \citet).
• IEEE venues use the cite package and IEEEtran (\cite).
• Never mix natbib and cite commands in one paper.

Review and verify the output

Treat the output as a first draft. The source states that drafted prose and results must be checked by the authors, and that the agent can still misstate findings. If you run the legacy citation mode (DBLP_BIBTEX set to false), some citations may carry [VERIFY] markers that you must resolve.

Read each section against your narrative and your data. Check every number, every claim and every reference before submitting.

• Compare quantitative results in the Experiments section against your own logs.
• Confirm each reference exists and matches the claim it supports.
• Search the output for [VERIFY] markers if you used legacy mode.
• Compile the project in LaTeX and fix any errors or overfull boxes.

Safety notes: scripts and network use

This skill does not run scripts. It does use the network: the default bibliography workflow fetches BibTeX data from DBLP and CrossRef. The draft content itself is produced by the agent's model, so the content of your paper is processed by that model as part of the session.

If your paper data must stay private, review what the agent sends and consider whether the network lookups are acceptable for your project. For Codex users with the reviewer overlay, the review step uses the claude-review MCP bridge and a Claude model.

• Scripts: none included.
• Network: yes, for DBLP/CrossRef bibliography lookups.
• Codex overlay: needs the claude-review MCP bridge and the skills-codex package.