# How to use Claude Code, Codex or ChatGPT to fix your paper after a review

> Turn reviewer comments into a prompt for Claude Code, Codex, Cursor or ChatGPT: what it needs, rules against invented results, and a template.

Updated 2026-11-15. 7 min read.

AI agents are good at the mechanical part of a revision: finding every passage a reviewer named, rewording claims, adding a baseline to the evaluation script, drafting the response letter. They are also good at inventing a result when they cannot find one. A good prompt gives the agent each problem with its exact location and fix, and forbids it from making up numbers or citations.

## Repo agent or chat assistant?

- Claude Code, Codex and Cursor work in a folder on your computer. Use them when your paper is in LaTeX or Markdown next to the code that produces its results: they can edit the source, change the evaluation scripts and run quick checks.
- ChatGPT or Claude in a chat window work from the paper you attach. Use them for a Word or PDF manuscript, or when you only need revised passages and an experiment plan.
- Either way, you check every change. The agent does the legwork; the results and the claims stay yours.

## What a revision prompt needs

1. **Each problem, located.** The exact sentence or section, the page, and what is wrong with it. "Improve the evaluation" gives a vague answer; "Table 2 reports one seed, p. 6" gives a fix.
2. **A suggested fix.** What the reviewer or you want instead, so the agent does not guess at the remedy.
3. **Text fixes and experiment fixes, apart.** Rewording a claim and running a new baseline are different jobs. Keep them in separate lists so the agent edits text directly and treats experiments as tasks.
4. **Ground rules.** Never invent results, numbers or citations. Leave a TODO where a result is missing. Soften a claim rather than strengthen the evidence on paper. For repo agents: do not commit, and stop before anything long or costly.
5. **What to hand back.** A change log by item, the TODOs left, the experiments still to run with their commands, and a draft response to the reviewers.

## A template you can adapt

Replace the bracketed parts with your own paper and the reviewer’s points, one numbered item per problem.

```text
You are helping me revise my research paper "[title]" and its experiments before I resubmit it. [Repo agents: This repository holds the paper source and the code that produces its results.] [Chat: I have attached the paper.]

## Ground rules
- Never invent results, numbers, data or citations. If a fix needs a result we do not have, describe the experiment and leave [TODO: ...] in the paper.
- When the evidence is weaker than the text says, soften the claim.
- Keep my voice and the paper's structure. Change what each item needs, not more.
- [Repo agents] Do not commit or push. Before anything long or costly, stop and tell me the command.

## Paper fixes
1. [MAJOR] [Short title] ([section], p. [n])
   Quote: "[the exact sentence]"
   Problem: [what is wrong]
   Suggested fix: [what to do]

## Experiment and analysis fixes
1. [CRITICAL] [Short title] ([section], p. [n])
   Quote: "[the exact sentence]"
   Problem: [what is wrong]
   Suggested fix: [what to run or report]

## Reviewer questions to answer
- [question]

## When you finish, give me
1. A change log: item number, what changed and where.
2. Every TODO you left and what it needs from me.
3. The experiments still to run, with commands, baselines and the result that would support or refute the claim.
4. A draft response-to-reviewers letter.
```

## Check the agent’s work

- Read the diff, not the summary. Agents describe changes more confidently than they make them.
- Search the revised paper for every number the agent touched and trace it to a script output or table.
- Look up every citation the agent added in Crossref or Google Scholar; reject any you cannot find.
- Run the experiments yourself and paste in the real results where the agent left TODOs.

## Getting the problems in the first place

The prompt is only as good as the list of problems in it. After a journal or conference review, copy the reviewers’ points in. Before submission, you need a reviewer of your own: a colleague, your supervisor, or an AI review that quotes your text and checks your references and numbers against the sources, so the agent is not fixing problems that do not exist.

## Do this faster with NoveltyAxis

Reviews your paper, checks every reference and number, and on paid plans writes this prompt for you from its own findings: each fix with its quote and page, the experiments to run, and the reviewers’ questions. See the [ai paper review](https://noveltyaxis.com/paper-review/index.md).

## Questions

### Can Claude Code or Codex revise a research paper?

Yes, when the paper source (LaTeX or Markdown) and its code are in the folder it works in. It can find the passages a reviewer named, edit them, change evaluation scripts and run quick checks. You still review every change and run the long experiments yourself.

### Can ChatGPT fix my paper after peer review?

It can draft revised passages, an experiment plan and a response letter from the paper you attach and the reviewer comments you paste. It cannot run your experiments, and it may invent a result or citation unless the prompt forbids it, so check everything it writes.

### How do I stop an AI agent inventing results?

Say so in the prompt, and give it a way out: tell it to leave a clearly marked TODO wherever a result is missing, and to soften claims rather than add evidence. Then trace every number in the revision back to an output you produced.

### Should I disclose that I used AI to revise my paper?

Follow your venue’s policy. Many publishers ask authors to disclose AI assistance with writing; none accept AI-generated results or citations presented as the authors’ own work.

### Does NoveltyAxis write this prompt for me?

Yes, on paid plans. After a paper review, select Fix with your AI agent and it writes one prompt from the review’s own findings, for Claude Code, Codex or Cursor, with a chat version for ChatGPT or Claude: every open fix with its quote and page, the experiments to run, numbers the body does not back up, references that failed to check, and the reviewers’ questions.

HTML version: https://noveltyaxis.com/guides/ai-prompt-to-revise-paper-after-review/
