Skip to content
Get packs

Guide · Agent skill

How to use the ML Paper Writing Assistant agent skill

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

The ml-paper-writing skill is a set of instructions and reference files that an AI agent loads when you ask for help with a machine learning or AI research paper. It is aimed at venues such as NeurIPS, ICML, ICLR, ACL, AAAI and COLM. After reading this guide you will know what the skill does, how to install it, how to phrase requests so the agent uses it, and what to check before anything goes to a conference.

ML Paper Writing AssistantOrchestra-Research/AI-Research-SKILLs · MIT

Quick start

  1. 1Download and unzip the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into .claude/skills/ in your project.
  2. 2Install the Python packages. Make sure semanticscholar, arxiv, habanero and requests are available in the environment the agent uses.
  3. 3Optionally add the Exa MCP server. If you want richer academic search, add the Exa MCP server using the command in this guide.
  4. 4Point the agent at your research repo. Open the project and ask the agent to explore the repository, results and existing citations before writing.
  5. 5Agree on the contribution. Let the agent state the main contribution and key results, then correct its framing where it is wrong.
  6. 6Review the first draft. Read the draft section by section, and ask the agent to list any placeholder citations it could not verify.
  7. 7Run the checklist. Ask the agent to check the draft against the relevant venue checklist, then verify the result yourself before submitting.

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

An agent skill is a folder containing a SKILL.md file and supporting reference files. When a request matches the skill's description, the agent reads these instructions and follows them. This skill is written for drafting ML and AI papers from an existing research repository.

Once it is active, the agent explores the repository, reads the results and any existing documentation, and finds papers already cited in the codebase. It then proposes a paper structure and produces a first draft for you to refine. The skill also includes LaTeX template notes for ACL, AAAI and COLM, conference checklist summaries, reviewer criteria, and a citation workflow that searches for and verifies references instead of writing them from memory.

  • Explores a research repo for key results, configs, and existing .bib files
  • Proposes a paper structure and drafts sections iteratively
  • Lists conference checklist requirements for NeurIPS, ICML, ICLR and ACL
  • Searches for and verifies citations, and marks unverifiable ones as placeholders

How do you install the ml-paper-writing skill?

Download the ZIP for this skill from this page. Then choose one of the two options below.

For a personal setup available in every project, unzip the folder into ~/.claude/skills/. For a setup limited to one project, unzip it into the .claude/skills/ folder inside that project.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

  • Personal install: unzip into ~/.claude/skills/
  • Project install: unzip into .claude/skills/ in your project folder
  • Claude apps: Settings → Capabilities → Skills, then upload

How do you trigger the skill with a natural request?

You do not need a special command. The agent decides to use the skill when your request matches its description, which covers drafting papers from research repos, structuring arguments, verifying citations, and preparing camera-ready submissions. Name the venue and point the agent at your project so it has the context it needs.

Requests that work well state the goal, the venue and the material. The skill also covers formatting for a conference and converting a paper to a different venue, so you can ask for those directly.

  • 'Draft a first paper from this repo for an ICML submission. Start with the results in results/.'
  • 'Propose a paper structure for this codebase and explain which claims you think are the main contribution.'
  • 'Check the references in my related work section and mark any I cannot verify.'
  • 'Convert this draft from the ICLR format to COLM and list what changes.'
  • 'Go through the NeurIPS checklist against my draft and tell me which items are missing.'

What do you need before you start?

The skill's SKILL.md lists its dependencies as the Python packages semanticscholar, arxiv, habanero and requests. It also says the skill runs scripts and uses the network, and that internet access is needed for literature search and API lookups.

Literature search can also use the optional Exa MCP server, which requires network access to mcp.exa.ai. The source gives this command for Claude Code: claude mcp add exa -- npx -y mcp-remote "https://mcp.exa.ai/mcp". Exa is optional. The skill's citation workflow still relies on search and verification APIs.

  • Python with the packages semanticscholar, arxiv, habanero and requests
  • Internet access for literature search and API lookups
  • Optional: Exa MCP server for academic search
claude mcp add exa -- npx -y mcp-remote "https://mcp.exa.ai/mcp"

How does the skill handle citations safely?

Citation handling is the skill's most important rule. It says never to generate BibTeX entries from memory. Instead the agent should search, verify the paper, and then fetch the BibTeX programmatically. If a reference cannot be verified, the agent should leave a clearly marked placeholder rather than guess.

This means you still have to check the output. The skill's own guidance requires human verification of claims and citations, and the agent should tell you which citations it could not confirm.

  • Search, verify, then fetch BibTeX. Do not accept BibTeX written from memory.
  • Unverifiable references are left as explicit placeholders, for example a \cite marked TODO for verification.
  • Read every reference in the final bibliography yourself before submission.
% EXPLICIT PLACEHOLDER - requires human verification \cite{PLACEHOLDER_author2024_verify_this} % TODO: Verify this citation exists

What should you check before submitting?

Conference checklists are part of the submission in the venues this skill covers, and the skill's reference notes say missing a required checklist can lead to desk rejection. Use the checklist reference to review your draft against each item, and check the venue's current rules yourself, because policies on AI-generated content vary.

Review the draft for claims that the results do not support. The reviewer guidance in the skill lists common reviewer concerns, which is a useful way to test your own argument before others do.

  • Check the venue's current checklist and formatting rules directly
  • Confirm that every claim in the abstract and introduction is supported by results
  • Read and verify every citation, including placeholders
  • Confirm your venue's policy on AI-generated text before you submit

Safety notes: scripts and network access

This skill runs scripts and uses the network. The agent may run shell commands such as listing files and searching the repository, and may call literature and citation APIs. Review what the agent proposes to run, especially in a repository that contains sensitive data.

Only install the Python packages you are comfortable with, and decide whether you want the optional Exa MCP server connected to your session.

  • Review commands before the agent runs them in your project
  • Be aware that citation lookups send paper titles or identifiers over the network
  • Keep the Exa MCP server optional if you do not want search traffic sent to mcp.exa.ai

What people use it for

  • Turning a finished experiment repository into a first draft for a NeurIPS or ICML submission
  • Restructuring an existing draft so the argument and contributions are easier to follow
  • Auditing the related work section and flagging references that cannot be confirmed
  • Converting a paper between ACL, AAAI and COLM formats using the template notes
  • Preparing a camera-ready version by checking it against venue formatting rules

Frequently asked questions

How do I install the ml-paper-writing skill?

Download the ZIP from this page and unzip it into ~/.claude/skills/ for all projects, or into .claude/skills/ inside one project. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.

Is the ml-paper-writing skill free to use?

The skill is published under the MIT licence, as stated in its metadata. Check the licence file in the source before redistributing it.

What does the ml-paper-writing skill require?

It lists the Python packages semanticscholar, arxiv, habanero and requests, and needs internet access for literature search. The Exa MCP server is optional.

Can the skill write a paper that is ready to submit?

It produces a first draft to refine with the researcher. The skill requires human verification of claims and citations, and you should review the venue's rules on AI-generated content before submitting.

Which venues does this skill cover, and what should I use for systems conferences?

It covers NeurIPS, ICML, ICLR, ACL, AAAI and COLM. For OSDI, NSDI, ASPLOS and SOSP, the skill points to a separate systems-paper-writing skill.

What happens if the agent cannot verify a citation?

The skill says to leave an explicit placeholder marked for verification rather than invent a reference. You then need to confirm or replace it yourself.

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

This skill suits ML researchers and graduate students who have a research codebase and want a structured first draft, help with references and a checklist review. It is not a fit for systems venues, which need the systems-paper-writing skill, or for anyone who wants to submit AI-written text without reviewing it and checking every citation.

Download the skill