Feedback Machines: Writing and editing research papers with generative AI

Claude Code and Cursor are changing the way we can get feedback on research

The classic way of getting feedback on your research paper was to ask someone to read it, hoping they would read past the introduction to give you a few notes. During the PhD, that person is usually the supervisor, and sending your work to them is usually quite stressful. “His palms are sweaty, knees weak, arms are heavy,” Eminem wrote, talking about a PhD student sending a draft to their supervisor. Other sources of feedback are conferences (which are sometimes sparsely attended) and referee reports (which are hard to interpret, and sometimes do not tell the whole story). The point is that it is difficult and time-consuming to get feedback on your work from others, and you cannot trust yourself to be an impartial judge of your work.

AI tools are changing this dynamic. It is now easy to get feedback on your own work, and if you are careful, most of us can become much better researchers by using AI tools. In this post, I will outline some ways I have been using Cursor and Claude Code to get quick feedback on my own work. I’m updating my website with tips for PhD students, including new resources on using AI tools, so please send along any guides or tips you have.

Integrating AI tools into the workflow

I recently started using Cursor to write and edit my papers. Cursor is an editor where you can write papers in LaTeX and write code. The main advantage (for me) of using Cursor is that I can very easily integrate AI tools into the work process in a natural way. I am not using it for ideas or coding, so this is more targeted toward improving paper writing.¹ There is a nice guide to Cursor and Claude Code here. Claude Code can write directly to your files and make edits to your code. Claude (and Cursor) can run your code, make changes to documents, and work through problems independently.

Why is this useful? Beyond the simple fact that it makes it slightly faster to edit papers, Claude will know more about your specific project. This is because you can give additional context and instructions to Claude, which helps it give you better feedback on your work. Importantly, you can write a guide to what the project is about, or even better, have Claude read the project and write you a guide. You can also give detailed instructions on how you want to paper to read. I realized this after reading scott cunningham’s Substack post on using Claude to make slides (which is great). Scott wrote about asking Claude to write down the tacit knowledge of what makes a great slide deck (you can read the results here). Reading through it, I recognize many of these principles from the presentation guides I linked on my website. I realized you can do the same thing for writing: ask Claude to give you a writing guide for your field, or simply put your favorite writing guide in the folder and let Claude read it. Once this document lives in your project folder (as a Markdown file named claude.md), you can get feedback on your work that ideally aligns with how you would like the paper to look. Here is what I got for writing. And here is a guide to best practices for Claude Code, including writing a claude.md document.

A concrete workflow example

To make this more concrete, here is a typical way I use Cursor after finishing a first full draft of a paper.

I start by asking for a referee-style report on the entire paper. The goal at this stage is not to implement every suggestion, but to get a high-level view of the paper’s perceived contribution, main weaknesses, and any claims that appear unsupported. This helps me see whether the paper I wrote is the paper I think I wrote.

“Write a referee report for this paper. Focus on constructive feedback, but make note of the strong and weak sides of the paper. Focus on finding claims that are not backed up. Imagine that you are a referee for a top 3 finance journal.”

The AI-reports I receive with this prompt focus a lot on the negatives. I once uploaded the paper that we recently published in RFS, and ChatGPT recommended major revisions. You still need to exercise some judgment yourself about which points to actually use in your own process.

Next, I do a section-by-section pass. For each major section, I ask Cursor to summarize the section and provide targeted feedback, and then to check whether it is consistent with the rest of the paper. This is often where inconsistencies show up: results that are framed differently across sections, or motivations that quietly drift over time.

Evaluate this section of the paper. Summarize the section first and then give feedback. Check if there are inconsistencies in the writing or parts that need to be clarified or explained better, and check if the results are consistent with the rest of the paper. 

I also ask for unsupported claims and missing robustness checks directly. This would give you a better sense of how you need to improve the paper. You don’t have to do all the suggestions, but it’s sometimes useful to get an idea of what you could do.

"Identify unsupported claims and missing robustness checks. Go through the paper and identify any empirical claims that are not directly supported by a table or figure. Also suggest robustness checks that referees might request based on the methodology used.”

After that, I run explicit consistency checks across the whole project: notation, variable definitions, reported magnitudes, and references to figures and tables. Because Cursor can see the entire folder, it is relatively good at catching mismatches between the main text, tables, and appendix that are easy to miss in a long project.

Examine if all figures and tables are referenced in the text, and see if they are well-explained. Also see if they are included in the right order. 

Only once the structure and internal consistency look right do I move to polishing. At that stage, I ask for professional editing focused on clarity and tone, and I tighten the introduction and conclusion so that they accurately reflect what the paper actually does, not what an earlier version used to do. Grammarly is probably better for this type of work, but it never hurts to try other programs.

Evaluate the entire paper. Act as a professional editor and make suggestions for writing, grammar and clarity. 

Once the paper is internally consistent, I turn to journal fit. If I do not yet have a target journal, I ask Cursor which journals the paper seems best suited for and why. If I already have a target in mind, I ask what aspects of the paper would likely need to change to better fit that journal. I do not treat these suggestions as requirements, but they are useful for clarifying what dimensions referees are likely to care about.

What journal should I target with this paper? What do I need to change or add to the paper to make it a good fit for [your target journal] 

Other useful things

Cursor is useful for many parts of the research process. For example, I once asked Cursor to write me a README file for the code. Since Cursor has access to all the project code, it could easily compile a description of my project. In another example, I needed to write some alt-text for figures. I simply pasted in the instructions for the alt-text and asked Cursor to update my figure captions. I also used Cursor with Claude Code/Codex to create a website with several different mortgage calculator tools (Want to know whether you should buy or rent? Click through!).

Limitations

Used carelessly, AI feedback tools can make a paper worse rather than better. My biggest concern is that the model smooths over conceptual problems rather than fixing them. Claude is good at making text sound coherent even when the underlying argument is weak, identification is unclear, or economic meaning is ambiguous. All chatbots are also way too nice to us, making our bad ideas sound good.

Finally, there is a temptation to accept edits wholesale because they “read better.” Be careful about accepting all edits. AI suggestions are candidate edits, not “always-accept” edits. It is also quite common to see edits where Claude makes up some story for the results that does not make sense, or where they hallucinate some result. I wrote before about learning to ignore (some) feedback. That applies doubly to AI feedback.

Final thoughts

Cursor and Claude Code have become essential parts of my writing workflow. They are very good at providing feedback on your work, which is very useful for improving as a researcher. My advice is to experiment. Try some prompts, modify them for your own needs, and see what works. Not every suggestion the AI makes will be useful, and you’ll still need to exercise judgment. But if it saves you even a few hours of tedious editing or helps you make your point better, that’s time you can spend on the parts of research that actually matter.

If you have prompts that work well for you, or thoughts on how to improve the ones above, I’d love to hear about them!

1. You can download Cursor for free, although there is a paid version as well. It’s built on VSCode, so you can also just use that, or you can use Claude Cowork or Google’s Antigravity. You can then incorporate either Claude Code, Codex, or use Cursor’s agent tab to ask questions.