Articles

Writing with AI?
Think like a journalist.
Give people a reason to read.

Two prompts, three AI models, and a draft of my own. What improved, what got lost, and how I used journalism and editing techniques to finish the article.

Anna Ostrovska · Published 5 November 2026

If you have something to say, don’t start with “once upon a time”. Find the main point, start there, build your argument, and cut until there’s nothing left to get rid of. Journalism and editing techniques, polished over decades, work just as well in marketing content.

They matter even more when you’re getting AI to write or edit for you. Below, I’ll use my own text to show how to do this, where LLMs can help, and where your judgement and editing are needed.

When I read content on company websites or personal blogs, I immediately notice the fluff and buzzwords before the author gets to the point. One of my personal favourites is some version of:

In today’s fast-paced digital landscape, staying ahead of the competition means embracing change and innovation.

You’ve seen openings like this. The article might even have strong points and useful insights further down. I’m patient, but even I’m quick to close the tab.

So I decided to test how AI would edit a rough draft with an opening like this, and whether the changes would actually make it better.

One draft, two prompts, three LLMs

I drafted the beginning of an article about making marketing content interesting and readable, then deliberately added a generic opening. It asked readers to think about the texts they enjoyed and explained that the world was saturated with content. The draft stopped where the practical advice would begin.

Then I put together two prompts. The General prompt asked AI to improve the text, make it clearer and more engaging, remove unnecessary wording, and keep it conversational. The Editorial prompt gave specific instructions based on journalism and editing techniques: what to improve, how to do it, what to keep, what not to add, and how to cut.

<draft>
Insert the original draft here.
</draft>

I tested both prompts with ChatGPT, Claude, and Gemini, starting a separate temporary or incognito chat for each run.

That same draft became the basis of the article you’re reading. So yes, you’re reading the experiment and its outcome at the same time. Isn’t it fun?

Did AI make my draft better?

The original English passage was 292 words. With the General prompt, two models made it longer and one made it shorter. With the Editorial prompt, all three cut it. Shorter didn’t necessarily mean better. Sometimes a few extra words helped.

Word counts exclude the main headline and the models’ explanations, but include internal subheadings. Percentages are rounded. ✅ Keep · ⚠️ Question or revise · ❌ Reject. These are my judgements about particular edits, not scores for a whole model.

Claude’s General response added 21 words, 11 of them in two new subheadings. Some of that extra length bought navigation. Counting words tells me what changed, not whether the change helped.

The opening alone produced three different decisions with the Editorial prompt: ChatGPT moved the generic paragraph, Claude kept it and defended that choice, and Gemini removed it. I would cut it. Asking readers to think about the texts they enjoy doesn’t add enough here to justify delaying the actual problem.

Some changes were smaller but still worth questioning. ChatGPT took:

“I’m patient, but even I’m quick to close the tab.”

And turned it into:

“And while I consider myself a patient reader, even I’m quick to close the tab.”

I didn’t need to consider myself patient.

For this draft, I found more edits I wanted to keep in the Editorial responses. But the General prompt also produced useful work, particularly Claude’s changes to the example’s placement and its signposts. I would still edit every version before publishing it. The models’ explanations needed checking too: a confident account of an edit doesn’t mean it was a good one.

Note: These were single responses to each prompt, not repeated tests or measurements of reader behaviour. The accompanying Lab Note includes the full outputs, model details, and a review with model names hidden that helped me refine these assessments.

Explore the full AI editing experiment →

Where I take over

At this point, I had six edits of an unfinished passage. I still had to decide what belonged in the article, choose its structure, and deliver the practical advice the draft had promised. Those decisions needed the same journalism and editing methods I’d asked AI to apply. Here’s how I used them, and how you can use them in your own writing.

Choose a structure that suits your objective

Start with what gives the reader a reason to continue: the finding, the result, the problem, or a specific situation that captures their attention. A research summary might open with its most significant finding. A case study might lead with the problem the client solved and the result they achieved.

A personal story can work too, provided readers can soon see where it’s leading. And “soon” is doing important work in that sentence. Don’t let the story bury the point.

As an editing rule of thumb, I aim to establish the substance in the first paragraph or two. If I’m opening with a story, I might give it three short paragraphs before explaining where I’m going. These aren’t fixed limits. But if I need much longer, I check whether I’m asking the reader to wait through background they don’t need yet.

In this article, I opened with my position and a promise to use my own text to show where AI helps and where editorial judgement is needed. Then came the familiar empty introduction and the story of the experiment. The opening gives you the substance; the experiment lets me demonstrate the editing decisions.

For more on the structures themselves, see Purdue’s guide to the inverted pyramid and Poynter’s explanation of the nut graf.

Then edit what you actually wrote

What is the one main idea the reader should take away?

“How to use AI in writing” is a topic. “AI can improve individual sentences while weakening the argument or changing the author’s voice” is a point you can explain and support. Write yours in one sentence, then check whether the draft delivers it.

Does each paragraph help make that point?

Explain an unfamiliar term before building an argument around it. Put the problem before the proposed solution. If a paragraph starts a separate discussion that you don’t need to explain your main idea, cut it or save it for another article. You can return to the main idea after an example and explain what that example demonstrates. Just don’t make the reader sit through the same explanation twice.

What supports the claim, and have you checked it?

Borrow another essential practice from journalism: fact-checking. Make clear whether a claim comes from your experience, someone else’s account, or research. Follow figures and quotations back to the original source, whether you found them through AI, Google, or another article. Check what was actually measured, when, and whether it supports what you’ve written. Where possible, cross-check against independent sources. Three articles repeating the same unsupported number don’t make it reliable.

While writing this article, I looked for recent data on how people read online. An AI answer gave me a familiar statistic: our attention span has dropped to eight seconds, shorter than a goldfish’s. I traced it. It spread from a 2015 Microsoft Canada report, which credited it to a site called Statistic Brain. The sources behind that point to a 2008 study of web browsing that measured how long visits lasted, not how long people can pay attention. A BBC investigation found no evidence for the goldfish comparison either.

Another widely quoted claim, that people read about 20% of the words on a page, does come from a real study: Nielsen Norman Group’s analysis, published in 2008 and based on 2005 data from 25 users. Useful, but nearly two decades old and an estimate rather than a measurement. I couldn’t find recent, robust research on how people read online, so I didn’t use either figure as evidence.

What can go?

Cut until there’s nothing left to get rid of. It’s a useful principle for copywriting generally, from a landing page to a presentation. Remove repeated explanations, unnecessary adjectives, and phrases such as “it is important to note that”. Replace “conduct an analysis of” with “analyse”. If you describe something as “innovative” or “seamless”, explain what it actually does or lose the adjective.

Check what your cuts change, though. “May increase revenue” and “increases revenue” make different claims. Use active voice when it makes responsibility clearer, but don’t invent an actor just to avoid a passive sentence.

Does it still sound like me?

I don’t need every joke or every favourite phrase, but I do need to recognise my meaning and my voice after the edit. Check that the specific observation, unusual detail, or bit of humour that made the piece interesting hasn’t disappeared.

In my draft, I wrote: “Then reading becomes painful, and nobody likes pain. Well, almost nobody.” ChatGPT changed it to: “Reading becomes a chore, and nobody likes a chore. Well, almost nobody.” The sentence still looks like a joke, but replacing “pain” with “a chore” loses the implication that made it one. That’s the kind of change I look for when I compare an edit with the original.

A better prompt still needs an editor

Would I use AI to edit this draft again? Yes. The Editorial prompt produced useful cuts and changes to the order of ideas. But I also had to reject edits that changed my meaning, flattened a joke, or made a claim stronger than I could justify. Writing detailed instructions didn’t remove the need to read the result critically.

And the draft was only the beginning of an article. It promised practical techniques without explaining them. I still had to build the argument, interpret the experiment, choose the examples, and give readers something they could use. That took several rounds of writing, rearranging, and cutting, including cuts to passages I’d already approved.

In my ongoing work, I also give AI editorial guidelines and examples of my writing, with notes on what to preserve and why. As we edit, I add useful corrections to that reference material.

We didn’t test that setup here. It would be a useful next comparison: what changes when the model gets the same editing brief alongside examples of my voice and previous editorial decisions?

Keep improving your prompt as you notice what the edits get wrong. If AI keeps removing your humour, show it what you want to preserve. If it strengthens claims, point out exactly where the meaning changed. Then check the next version. A better brief helps, but you still need to know what a better text looks like.

Explore the full AI editing experiment →