ChatGPT for Writers: How I Use It Without Letting It Write for Me
I use ChatGPT for writing almost every day, which makes what I am about to say slightly inconvenient: I rarely ask it to write for me.
That isn’t a protest against AI. I use it too much for that. ChatGPT has become part of how I research, structure content, analyse search results, work with localisation, question arguments and edit what I have written. It can make parts of the process considerably faster, and occasionally it notices something I have missed.
But there is a difference between using ChatGPT while writing and asking ChatGPT to be the writer. The longer I use it, the more important that difference has become to me.
If you search for advice on using ChatGPT for writing, you will find an impressive number of prompts. Prompts for blog posts, prompts for novels, prompts for headlines, prompts for rewriting your prose in the style of somebody whose lawyers may eventually become interested. There is nothing inherently wrong with prompts. I use them too. But I think we sometimes focus so much on finding the perfect instruction that we skip the more important question.
What exactly should ChatGPT be doing in the writing process?
I don’t start with “write me an article”
ChatGPT can write an article from a prompt. That may be useful when the objective is simply to produce a piece of functional text quickly. But when the thinking matters, I find the blank page more useful than it looks.
Writing a first draft is often where I discover whether I actually have an argument. Something that sounded perfectly sensible in my head becomes less convincing when I have to explain it. An example changes what I thought the section was about. I write three paragraphs and discover that the second one contains the idea I should have started with.
None of this is particularly efficient.
It is also part of writing.
If I ask ChatGPT for the article before I have worked out what I want to say, it makes many of those early decisions for me. It decides on a likely structure, chooses which points deserve attention and produces the language in which I first encounter the argument. I can change everything afterwards, but I am now reacting to an answer rather than finding one.
That is why my better uses of ChatGPT usually begin with something of my own: an observation, a question, research, a draft or an argument that isn’t working yet. I’ve written more broadly about how I use AI for writing, including the parts of the process I deliberately keep for myself. ChatGPT is one tool within that larger question.
Then things become interesting.
I use ChatGPT to make research faster, not to replace sources
Research is one of the most useful applications of ChatGPT for writers. A broad subject that would once have taken considerable time to map can become navigable very quickly. I can explore terminology, competing positions, related concepts and questions I hadn’t considered. When I work on SEO content, AI can also help me understand the territory around a search query and identify aspects of the search intent I may have overlooked.
What I don’t want is for the process to end there.
ChatGPT is not the source simply because it gave me the information. If a claim matters to an article, I want to know where it came from. If a study supposedly found something useful, I want the study. If an organisation has changed a policy, I want its current documentation. And if a number looks suspiciously convenient, I become particularly interested in finding out where it came from.
This sounds obvious. It becomes less obvious when an AI answer is fluent, detailed and confident.
The speed creates its own temptation. When a model has summarised something perfectly well, opening five original sources suddenly feels like administrative work. Unfortunately, facts remain stubbornly uninterested in my productivity.
So I use ChatGPT to find paths through research. I don’t want it to decide where the path ends.
I ask it to disagree with me
This may be the most valuable thing ChatGPT does for my writing.
Once I have an argument, I can ask the model to attack it.
Not to produce three polite “areas for improvement”, but to look for the assumption I haven’t justified. What would someone who disagrees with this say? What evidence would weaken the argument? Am I presenting a correlation as a cause? Is there an obvious counterexample? Have I defined a term in a way that conveniently makes my conclusion true?
This became particularly useful while thinking about AI and writing itself.
For example, I argue that as AI makes it easier to produce language, judgment, selection, and responsibility become more important parts of the writer’s role. That sounds plausible to me. Which is precisely why it needs trouble.
Why wouldn’t AI also become better at judgment? Am I simply redefining “real writing” around the parts of my profession I hope to keep? If an AI system proposes an argument, finds the evidence and produces the prose, how much human selection is enough before calling the result authorship becomes rather generous? That question also sits at the centre of whether AI will replace writers. The issue isn't simply which tasks a machine can perform, but what happens to the writer's role when it can perform more of them.
Those questions don’t destroy the argument. They make me decide whether I can defend it.
There is a strange temptation when working with conversational AI to treat agreement as evidence. It isn’t. A model that tells me my idea is insightful may simply be very good at being pleasant.
I don’t particularly need that from an editor.
ChatGPT is useful for structure once I know what I am structuring
ChatGPT is extraordinarily good at outlines. Give it a subject and within seconds you can have an introduction, several sensible sections, common mistakes, best practices and a conclusion.
This is useful.
It is also how a lot of the internet ended up looking as if it attends the same meetings.
I use AI for structure, but preferably after I have established the central thought. At that point I can ask whether the order makes sense, whether I am answering the reader’s main question too late, whether two sections are doing the same job or whether the article contains a detour that interests me considerably more than it is likely to interest anyone else.
For search-driven content, I can also test whether the structure actually serves the search intent. If someone arrives because they searched for “ChatGPT for writers”, they reasonably expect to learn how a writer can use ChatGPT. An eight-hundred-word philosophical introduction to authorship before answering that question may be fascinating to me and slightly less fascinating to them.
SEO matters here.
But there is a difference between allowing search intent to shape what an article needs to answer and allowing a keyword to dictate how I write every paragraph.
The first helps the reader.
The second tends to produce articles nobody would voluntarily read if Google hadn’t introduced them.
I prefer asking for criticism to asking for rewrites
Editing is another area where ChatGPT is genuinely useful, but the prompt matters less than the role I give it.
“Improve this paragraph” gives the model enormous freedom. It can change the rhythm, vocabulary, emphasis and sometimes even the thought itself. The result may sound smoother. It may also sound less like me.
So I increasingly prefer diagnostic questions.
Where does this become unclear? Which sentence repeats an idea? Where does the argument jump too quickly? Which claim needs evidence? Does anything sound generic compared with the rest of the text? Is there a paragraph that could disappear without changing what the article says?
Those questions keep an important distinction intact.
ChatGPT identifies a problem.
I decide what to do about it.
Sometimes I take its suggestion. Sometimes I don’t. Sometimes it identifies the right problem and proposes exactly the wrong solution, which is still useful because I now know where to look.
This is also one of the simplest ways I have found to use ChatGPT without gradually losing my writing voice. I don’t need the model to rewrite everything in “my style”. I need it to help me see the text more clearly.
ChatGPT can imitate your voice. That doesn’t mean it has found it.
There are increasingly sophisticated ways to teach ChatGPT how you write. Give it enough examples and it can identify recurring sentence structures, vocabulary, tone and stylistic habits. It can then reproduce some of them surprisingly well.
That can be useful. I have examples of my own writing for exactly this reason when working with AI.
But I don’t think writing voice can be reduced to style.
Voice also comes from what a writer notices. It comes from which contradiction is interesting enough to pursue, which example feels wrong, what deserves a joke and what doesn’t, where certainty would be easier but doubt is more accurate. Two writers can use almost identical vocabulary and still produce completely different pieces because they pay attention to different things.
ChatGPT can learn patterns from previous texts.
It cannot retroactively have the experience that made me notice something in the first place.
That distinction matters to me because “write this in my voice” can otherwise become a strange shortcut. The result may resemble something I would write while quietly removing the part where I actually had to think about what I wanted to say.
I would rather use AI to protect the conditions in which voice emerges than ask it to manufacture the surface afterwards.
Prompts are useful, but I don’t think writers need a prompt library the size of a small novel
There are prompts I return to conceptually because they give ChatGPT a useful role. They are usually much less elaborate than the giant prompt templates circulating online.
I want to know what I am missing.
I want the strongest counterargument.
I want to know where the reader may become confused.
I want claims that need verification identified.
I want to know whether the structure answers the question someone actually searched for.
I may want alternative ways of expressing something when I know what I mean but the sentence refuses to cooperate.
The exact wording changes depending on the work. That is why I am reluctant to pretend there are ten magical prompts every writer should copy.
A useful prompt is not an incantation.
It is a clear description of the job you want the model to perform.
For writers, I think that becomes particularly important because the easiest prompt is also the one that delegates the most:
“Write this for me.”
Sometimes that is exactly the right prompt.
But it should probably be a decision rather than a default.
Where I do let ChatGPT produce text
There is no purity test here.
I use AI for localisation and alternative formulations. I may want variants of functional copy. I may want to see several ways of expressing a point. I may use generated material as something to react to. For work where the objective is clarity and efficiency rather than discovering a thought through the act of writing, having AI produce text can be entirely sensible.
The distinction isn’t “AI may never write a sentence”.
That would be difficult to reconcile with actually using AI.
The distinction I find more useful is whether I understand which decision I am delegating along with the text. I created the Human Decision Map to explore exactly that distinction. It lets you look at where AI contributes to your writing process and where you are still making the decisions that shape the finished work.
If I ask for ten headline alternatives, AI is generating options and I am selecting among them. If I ask it to translate or localise something, I still need to decide whether the result carries the intended meaning and works in its new context. If I ask for a rewrite, I need to notice when clarity has improved at the cost of something else.
This is the practical side of what I mean by an AI Writer.
I use that term for a human writer who deliberately integrates AI into the writing process while retaining authorial judgment, creative direction and responsibility for the final work.
The important word isn’t AI.
It is writer.
ChatGPT becomes more useful when I give it less authority
That sounds slightly paradoxical. The more capable the tool becomes, the more tempting it is to give it the entire task.
Research this. Decide the angle. Create the outline. Write the article. Optimise it. Edit it. Give me the finished version.
It can do that.
And sometimes that is useful.
But for the writing that matters to me, I get more value from separating those stages and deciding where I actually need help. ChatGPT may be a researcher for ten minutes, an irritating opponent for the next twenty and an editor later. It doesn’t have to become the author simply because it is capable of producing words.
This way of working has also changed how I think about creative work beyond text. With Kuh Alma, for example, generative AI lets me move from writing into images, music and video and generate possibilities for a fictional world much faster than I could produce them alone. But the same problem returns in another form. I explored this more concretely when I used generative AI to turn Kuh Alma's written world into a film trailer. It was an experiment in storytelling, but also in deciding which generated possibilities actually belonged to the character.
AI can generate possibilities.
Someone still has to decide what belongs.
I use the term author-led storyworld for a storyworld in which an identifiable human author retains final creative authority over its canon, direction and meaning, even when AI contributes substantially to its development, expansion or production.
That is a bigger subject than ChatGPT for writers. But it comes from the same experience.
The interesting part of generative AI isn’t only that it can make more things.
It is what happens to the person making the decisions.
So how should writers use ChatGPT?
I don’t think there is one correct workflow. A journalist, novelist, SEO writer, screenwriter and technical writer have different obligations and different reasons for writing. Someone trying to finish an email before lunch is solving a different problem from someone trying to find the ending of a short story.
But I have found one question more useful than any collection of prompts:
What do I want ChatGPT to help me do, and what do I still want to decide myself?
That question works for research. It works for outlining. It works for editing, rewriting and creative exploration. It also makes it easier to notice when convenience has quietly turned into delegation.
ChatGPT has made parts of my writing much faster. It has made research easier to enter, arguments easier to test and drafts easier to examine from another angle. I don’t see much value in pretending I would be a better writer by refusing to use it.
But I also don’t want to become the person who enters a prompt, receives a text and spends the rest of the process making it look as if something more interesting happened.
The blank page is less lonely now.
I am not entirely sure it should become empty of me.
Frequently Asked Questions About ChatGPT for Writers
What is the best way to use ChatGPT for writing?
I find ChatGPT most useful when I give it a specific role rather than an entire article to produce. I use it to explore research questions, challenge arguments, examine structure and identify weaknesses in drafts. It can also help with editing and alternative formulations. But I prefer to bring my own idea or text into the conversation first. That way, ChatGPT helps me develop the work rather than deciding what the work should be.
Can ChatGPT help improve my writing without rewriting it?
Yes. In fact, that's one of my favourite ways to use it. Instead of asking ChatGPT to improve a paragraph, I ask where the argument becomes unclear, which sentences repeat an idea or where a claim needs evidence. Sometimes it identifies the right problem and suggests the wrong solution. That's still useful. I get to see what needs attention without automatically adopting somebody else's version of my sentence. Even when that somebody is a machine.
How do I use ChatGPT for research without relying on inaccurate information?
I use ChatGPT to explore a subject, understand terminology, identify competing arguments and find questions worth investigating. But I don't treat its answers as verified sources. If a claim matters, I check the original study, publication or documentation. ChatGPT can make research considerably faster, especially at the beginning. It doesn't remove the responsibility to find out whether something is actually true.
About the Author
Carsten Jan Weichelt is a professional content writer and strategist who works with AI in research, SEO, content development and editing. He also writes fiction and experiments with generative AI in creative storytelling, including his project Kuh Alma.
He uses ChatGPT as part of his professional writing process, particularly to explore research, challenge arguments, examine structure and improve drafts. His interest lies not only in what AI can produce, but in how writers can use it without giving up their own judgment, creative direction and responsibility.