Style Guides for AI Writing: Getting a Specific Voice

Even though language models have “language” in the name, they can still disappoint when it comes to writing—especially the kind of writing we’d call distinctly human. They’ve gotten much better, but they’re not exactly known for producing long, compelling prose or short stories that feel like they came from a single mind with a point of view.

There are a couple reasons for this. Part of it is that great creative writing simply hasn’t been the main focus for most AI labs. But the deeper reason is the core way these models work.

If you ask an AI to write something, you have to think about how it was trained. It learned language by reading an enormous amount of text and generalizing from it. So if you ask it to write a reference article, it draws on millions of Wikipedia articles and gives you something like an average of them. If you ask for a blog post, it’s using rules that come from the average blog posts it’s seen (even if it’s implicitly selecting “better” ones).

And the problem is: there’s a world of difference between a feature writer, an editorial writer, and a marketing copywriter. But the model has to learn from all of those at once, and then produce general rules that apply across all of them.

That’s why, when you prompt a model without much guidance, you often get text that’s readable, correctly formatted, and grammatically clean—but it doesn’t quite flow like it came from a single author. It feels like it was written by a committee. No terrible choices, but not many inspired ones either. That’s the crux of why these models often aren’t as good at writing as people expect them to be.

The other major issue shows up when you ask for something long—say, a 10,000-word story, let alone a 100,000-word novel. I think these models can write really good 200–300 word short stories. But longer fiction isn’t just “more words.” It needs structure underneath: story arc, character development, act structure, pacing, setups and payoffs. Those are patterns that often have to be explicitly explained to us before we even notice them.

If I read a thousand novels, I might still not fully realize how consistent the structure is across them—because the words are all different. That’s the trick. The patterns are there, but they’re hidden inside the prose.

The good news is that models are very good at breaking things down. And one of the best ways to get a model to write better is to have it uncover the pattern first, then write from that pattern.

Here’s what I mean. If I want a model to write like Sir Arthur Conan Doyle, I could just say, “Write me a story in the style of Sir Arthur Conan Doyle.” And sure—it’s going to mention Sherlock, Watson, Victorian England, and it’ll use some period conventions and familiar phrases.

But it still might not actually write like Conan Doyle. There are subtler patterns it might not automatically lock onto: the way he opens a paragraph, the way he closes it, the rhythm of his observations, what he does and doesn’t describe, the kind of judgment the narrator makes.

So the easy workaround is: don’t start by asking it to imitate. Start by asking it to explain.

Tell the model: “Describe how Sir Arthur Conan Doyle writes. Give me an examination of his style. Show examples. Critique it.” Let it generate a document that makes the style explicit: vocabulary choices, what he would never do, what he tends to emphasize, how scenes are framed, how information is revealed.

Then you take that document and say: “Now write using this style. Here is the style guide.”

In practice, you usually get much closer to what you actually wanted.

You can apply the same idea to your own voice. Give the model a bunch of your writing and ask: “What are my patterns?” It might notice your word choice, the kinds of examples you use, whether you avoid vulgarity or lean into it, how you structure arguments, how you transition between ideas. Once it states those patterns explicitly, you’ve basically created a reusable prompt—a style guide for yourself.

There’s still nuance beyond that, obviously. But the basic approach is simple: have the model explain the style first, then use that explanation as part of the prompt. Because if you only say “use this style,” the model often won’t do the work of locating all the relevant patterns it needs to start from in order to really deliver.

And since I’m a novelist and I work in AI, I get asked this all the time: do I have AI write my books?

Obviously not.

If I were a pianist or guitarist—and I was into robots—would you ask me if I used a robot to play the piano for me? I love writing. And yes, I’m confident these models will get better. I think they’ll be able to surpass even the best writers of today. But that doesn’t mean I’m going to enjoy the process any less than I do now.