How to use the Research Paper Writing agent skill

What is an agent skill and what does this one do?

An agent skill is a folder containing a SKILL.md file and supporting reference files. When an agent decides a task matches the skill's description, it loads the relevant guidance and follows it. You do not need to call it by a command; you describe what you need in ordinary language.

This skill is built for academic papers in the ML, CV and NLP style. It routes your request to the matching guide: Abstract, Introduction, Related Work, Method, Experiments, Conclusion, a five-dimension adversarial review, paragraph flow checks, or LaTeX-to-PDF compilation. It loads only the guide it needs for the task.

• Polishes a single paragraph and explains briefly what changed.
• Drafts or rewrites a section, with a short outline before the prose and a claim-evidence map for Abstract and Introduction.
• Runs a pre-submission review and returns a prioritized fix list, rejection-level risks first.
• Edits .tex or .md files in place while keeping citations, labels, comments and math environments unchanged.
• Turns bullet points or Chinese notes into an English section, after collecting the facts it needs.

Install the skill from the ZIP

Download the ZIP of the research-paper-writing skill from this page. Then choose one of the two install locations below. The folder you unzip should contain SKILL.md.

For a personal install that works in every project, unzip the skill into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder so it applies only there. Start a new session afterwards so the agent can see the skill.

If you use the Claude apps instead of a local folder, open Settings, go to Capabilities, then Skills, and upload the skill there.

• Personal: ~/.claude/skills/ (available across projects).
• Project: .claude/skills/ inside your project folder (available only there).
• Claude apps: Settings → Capabilities → Skills, then upload.

Trigger the skill with natural requests

You do not need special syntax. The skill's description lists the situations it covers, so ask for what you want in plain words. Mentioning your paper, draft, camera-ready version, or a target venue such as CVPR, ICCV, NeurIPS, ICLR or ACL helps the agent recognise the task. Mentioning a .tex file being edited for a paper also works.

For files, point the agent at the paper folder. The skill tells the agent to locate sections with grep, for example searching for \section or \begin{abstract}, and then edit them in place.

• "Polish this paragraph. Does it flow?"
• "Rewrite my Abstract using the challenge-to-contribution structure."
• "Here are my bullet points for the Method section. Turn them into publication-quality English."
• "Review my paper as a skeptical reviewer before I submit to NeurIPS."
• "Compile this paper to PDF; the build is failing."
• "这段通顺吗？" or "编译PDF" also trigger the skill, according to its description.

What to expect from the output

The skill is designed to avoid making things up. If an experimental number, citation, baseline name or dataset statistic is missing, the agent should insert a visible placeholder such as [XX.X] or [CITE: sparse-view NeRF methods] and tell you what to fill in. If a claim in your Abstract or Introduction has no backing in the paper, the agent should propose a weaker version and explain why.

The agent may ask up to three focused questions if key facts are missing, such as the exact problem, the contribution, why the method works, and the strongest results. For a quick polish you should receive the revised text plus one or two lines on the changes, not an outline.

• Placeholders appear where facts are missing; check every one before submission.
• Lines starting with ⚠ mark sentences where the agent had to choose an interpretation.
• Section drafts include a claim-evidence map: Claim, Evidence, Status (supported, needs evidence or weakened).
• If the draft already reveals author identity, the agent is instructed to warn you, since the skill respects double-blind review.

Safety notes: scripts and network

According to the skill's description, it does not run scripts and does not use the network. It works through the instructions in its reference files and by reading and editing the paper files you give it access to.

Compilation to PDF depends on a LaTeX toolchain on your machine. The skill's PDF guide covers engine choice and fixing failures, but the source does not name specific tools, so you need to have a working LaTeX setup. Review any file edits with your normal version control before accepting them.

• Runs scripts: no.
• Uses network: no.
• Needs access to your .tex or .md paper files for in-place editing.
• Needs a LaTeX toolchain only if you ask for PDF compilation.

Tips for better results

Give the agent the facts first: the exact problem, why earlier methods fail, what your contribution is, and your strongest numbers. The skill is built around correct facts arranged well, so better inputs produce a better draft.

Ask for one section at a time, and keep the terminology in your paper consistent. The guidance asks the agent to avoid swapping synonyms for a key concept.

• Paste your real results tables or numbers so the agent does not need placeholders.
• Use the review request before submission to get a prioritized list of risks.
• For camera-ready or rebuttal-facing edits, say so explicitly so the agent knows the context.
• Check that every citation key, label and equation survived the edit.