My last 2 posts have dealt with 1) the use of Claude to complete a peer review, and 2) how journals could include this process in their workflow.
It follows that authors should be using LLMs as well. There are the obvious use cases; rephrasing passages, summarising, expanding, correcting, and so on. However, I think we can go further and use language models for a wider range of more cognitively demanding tasks. For example, exploring connections between seemingly unrelated ideas or knowledge domains, or ensuring that the writing is aligned with the journal’s aim and audience.
In this example, I’m suggesting Claude and Lex but there are other AI-supported writing tools that are increasingly available. And I’m also working from the perspective of an early-career researcher, or postgraduate student.
I think you’ll probably need to experiment with different prompts and maybe even different models, so see what gives you the best results. The suggestions below are my first attempts to prompt Claude, and I was surprised at the quality of the responses. I feel like I could definitely be using this process for academic writing.
A workflow for using LLMs to help write a first draft of an academic paper might look something like this:
- Open Claude and enter the prompt: “I am an early career researcher, starting out on my journey of learning how to write academic articles for publication. You are a much more experienced researcher, with many years of writing for publication. I would like your help to structure my first paper. We are going to go through the process of outlining an article together, and you will give me guidance on how to do that. I will begin by giving you my research question and the aim of this article. I would like you to suggest an outline of what an introduction to this article might need to cover. Do you understand?”
- Claude will then respond to confirm the process.
- The next prompt might look something like this: “My PhD thesis tried to answer the question, [insert research question here]. The aim of the article I’d like to draft today is [insert article aim here]. Please can you give me an outline of the kinds of things a good academic article on this topic should include in the Introduction.”
- Claude will then give you a bulleted list of 5-10 items, suggesting an outline for the Introduction section.
Copy the outline and into a new document in Lex.
- Don’t bother giving it a title. Just paste the outline and tidy it up i.e. remove the intro and concluding text that Claude generated. And remove the bullets.
- Edit the relevant detail that Claude generated, to make sure it’s aligned with your thinking; remember that Claude has no idea what’s in your thesis, and your article needs to be grounded in that reality. This means removing items that aren’t relevant in the context of this article, and maybe adding a few brief sentences for anything that Claude didn’t include, but that is important.
- At the end of each sentence, simply type +++.
- Lex will now start generating content, using the initial sentence as a prompt. I find that it generates too much and seems to use all the content on the page to inform its response. You can stop the generation process at the end of the paragraph, or let it run and delete what isn’t necessary.
- Once you’ve done this for all the sentences, go to the star icon at the top right of the page, and click it to generate some titles. Copy and paste one you like, into the title section.
Use a version of this process to move through the rest of the article. The next section will be the Methods. Go back to Claude and use the prompt to briefly explain the process you used to conduct the research. Ask Claude to describe what a rigorous and systematic approach to answering the research question might look like i.e. give it your design, sample, data collection and analysis, etc.
Here is an example prompt that generated a reasonable outline for the Methods section: “In order to answer this research question, I interviewed a diverse sample of physiotherapy students across the cohort. I did the interviews in person, having obtained informed consent from everyone. Please give me an outline describing a systematic approach to conducting interviews, including the process for recording, transcribing, and analysing the qualitative data. Note that I used Braun and Clarke’s 6 step process for the data analysis.”
Follow the same process with Lex to expand on the outline from Claude, again making sure to edit the content so that it reflects what you actually did.
After this, I think that it’s probably reasonable to say that you should write the Results section yourself. However, if you haven’t analysed the data (or you want to get a 3rd perspective on your analysis), you could submit your anonymised transcript to Claude and ask it for an analysis. I won’t go into detail with that, as that’s the subject for another post.
For the Discussion section, I’d go back to Claude, and give it a very high-level summary of the findings of the study, and ask it to generate an outline of what the discussion should include, making sure to align the outline with the aim of the paper, and the findings that you just gave it.
Here is an example prompt that I gave me an outline for the Discussion that I thought was remarkably good: “I identified the following themes after analysing the data; 1) students spent less time preparing for assessment tasks immediately before the assessment, but more time overall, 2) they felt less stressed for assessments, 3) they felt more confident in their understanding. Please can you relate these findings to some of the points you made in in the introduction. I’d also like you to explore the connections between these findings, and any relevant learning theory you’re aware of.”
And again, take the output from Claude and use Lex to expand on it. And again, make sure that you edit the generated outputs to make sure it’s all aligned with what was actually done. Remember that LLMs are always making things up, and you need to ensure that what’s being generated is related to what you did.
Finally, ask Claude to write a short conclusion that summarises the process, highlights the main findings, and suggests a meaningful and realistic impact for the sample population.
Here is an example prompt for the Conclusion: “That’s a great outline for an article on this topic. Please give me a brief summary of what was done, what was found, and conclude with the implications for physiotherapy educators.”