How to write a paper with LLMs
It takes a great prompt!
In my post yesterday, I showed something I think is of great importance, namely how to write with AI a meaningful research paper, with references and diagrams for better illustrations of key concepts. Andrej Karpathy has posted yesterday on X how go about doing such a thing, namely to pay attention to context as well as prompt engineering. I have a great deal of respect for him and I do not disagree with his points, context is very important in all kinds of things, such as AI and quantum computers, to name a few.
But we all want great results as quickly as possible with least amount of time and effort, and great prompts are indeed key, so let us go into more details what I mean by this observation. In my post yesterday I noted how my post two days ago was basically the prompt for that paper. I used it verbatim as a prompt to Claude 4 but I could have shortened it too, as I will describe in a moment.
But first, let me explain why I used Claude 4, instead of ChatGPT-4.5 which I do use a lot. It turns out Claude 4 is much better at writing very high quality content in a SINGLE SHOT, once you properly motivate and prompt it. Note that you have to be firm and direct with your interactions, namely not to allow it to give you mainstream responses full of hedging and vague responses full of “yes, but …” qualifications. You may think of a politician or a mainstream article or a video where you ask yourself what they actually wanted to say and can not figure it out. It is not you, they basically said NOTHING and created an appearance of some meaningful analysis or discussion.
Do NOT allow LLMs to behave like that, call them OUT right away, they are VERY responsive to such behaviors, as they cannot get offended by definition, and will interpret your strong and categorical opinions as guidance what to say.
Now let us go back to Claude 4, it can do even 20+ pages of very high quality content if you give it the RIGHT PROMPT. So what was the right prompt about in my post two days ago? It was about several key points:
Reasoning is mechanistic and can be done by machines in automated fashion. Indeed such automatic enumerations are key parts of all kinds of formal proofs in computer science and theory of computing.
Kurt Gödel used such a scheme, namely Gödel numbering in his seminal works on Incompleteness theorems to show the limits of reasoning for BOTH humans as well as machines. Note that ALL of his work is fully applicable to AI, with NO exceptions!
Formal propositions, proofs, logical formulae and indded entire systems can bee mapped to natural numbers in maps called bijections, that are basically one-to-one (as well as onto for those paying close attention). This seemingly innocent assumption is actually absolutely a crucial one.
It is the above three points that are immediately picked up by LLMs to understand what you mean and are trying to express. they are ubelievably good in such interpretations. The key is for us humans to give clear and succinct prompts based on such essential premises and LLMs will amaze you by what they are capable of crafting. Note that YOU are key in creating such crafts but so are LLMs which act as super high-powered assistants to such a high level that I personally view them as valid co-authors.
Let me give you a kind of meta-example of what I jut said, here is a research paper created by Claude 4 with the above as a prompt:
You can play with it as much as you want and , of course, you can try out all kinds of strong prompts you come up yourself, the key is to get one yourself :)