I’ve been writing for a living for eleven years, and I’ve had a ChatGPT tab open for most of the last two. Not because it writes better than me. It doesn’t. But it does something I’m terrible at: it produces a fast, mediocre first draft without complaining, and that draft is usually enough to get me unstuck.
That’s the honest version of ChatGPT for writing. It’s not a ghostwriter and it’s not a replacement. It’s closer to a very fast, very confident assistant who has read a lot and understands very little. Use it for what that describes, and you’ll save real hours. Use it for anything else and you’ll produce 800 words of smooth, hollow mush that everyone recognises instantly.
Where ChatGPT actually earns its keep
After a couple of years of daily use, I can divide the useful jobs into three buckets.
1. Getting from blank page to something
The hardest part of any piece is sentence one. Ask ChatGPT for five different opening paragraphs on a topic and you’ll hate four of them. The fifth will be wrong in an interesting way, and you’ll fix it in thirty seconds. You’ve now started writing. That’s worth more than the output itself.
2. Structural work
Give it your messy notes and ask it to propose three outlines with different logic. One might be chronological, another problem-solution, another by audience type. You will almost always pick a hybrid of two, and you’ll get there in four minutes instead of forty.
3. The grind tasks nobody enjoys
- Turning a 900-word draft into a 180-word meta description
- Writing fifteen subject line variations for the same email
- Rewriting a paragraph three ways for different reading levels
- Cutting a 1,400-word piece by 20% without losing the argument
- Naming things — products, sections, newsletter issues
None of these require taste. They require volume, and volume is exactly what a language model is good at.
The parts it quietly gets wrong
Here’s the thing about ChatGPT for writing that catches people out: it fails in ways that look like success. A flat sentence is easier to spot than a confident sentence that’s subtly false.
It invents specifics. Ask for a statistic and you’ll get one. Ask where it came from and you’ll often get a citation that doesn’t exist. I’ve had it fabricate a plausible-sounding 2019 study with a real journal name and a fake author. Every number, name, date and quote needs checking against a source you can open.
It defaults to a voice. Left alone, ChatGPT writes like a corporate brochure. Lots of “moreover”, lots of three-item lists, lots of sentences that hedge in both directions and commit to nothing. Ask it what it thinks and it’ll tell you there are many perspectives worth considering. You already knew that.
It can’t tell you what to leave out. It has no idea which detail is the interesting one. That judgement is the whole job, and it’s still yours.
Prompting for writing is different from prompting for answers
Most people prompt ChatGPT like they’re searching Google: a topic, a word count, a polite please. Then they wonder why the output reads like a Wikipedia summary with extra adjectives.
The fix is to give the model a role, a reader, a constraint and a sample of your actual voice. Something like: “You’re writing a 400-word section for a newsletter read by freelance designers who are skint and sceptical. No exclamation marks. No bullet points. My last three paragraphs are pasted below — match that rhythm.”
That single change does more than any list of “magic prompts”. If you want a deeper set of examples, there’s a solid collection of ChatGPT prompts that produce real work instead of polite filler worth working through — the ones built around constraints rather than topics are the ones that stick.
Feed it your own writing
Paste in 300 words of something you’ve written and liked. Ask it to describe the style in concrete terms: average sentence length, use of contractions, how often you use first person. Then ask it to write in that style. It won’t nail it, but it’ll get close enough that editing becomes rewriting rather than total replacement.
A workflow that keeps you in charge
This is the routine I’ve settled into, and it holds up across blog posts, newsletters and landing pages.
Step one: I write the thesis in one sentence, by hand. If I can’t, the piece isn’t ready for help yet.
Step two: ChatGPT produces an outline and three opening options. I take the worst opening and fix it, because a bad opening I understand is easier to improve than a good one I didn’t write.
Step three: I write the middle sections myself, using its outline as scaffolding and ignoring it wherever it’s wrong.
Step four: Back to ChatGPT for the boring passes — tightening, varying sentence length, drafting headline variants.
Step five: A read-aloud edit. Anything I stumble over goes. This catches AI-flavoured phrasing better than any tool, because the model writes sentences that look fine and breathe badly.
Spotting — and killing — the AI voice
Readers are getting good at this. Certain patterns give a draft away within a paragraph, and once a reader clocks it, trust drops and rarely comes back.
Watch for triads (three adjectives, three examples, three clauses), sentences that start with a participle clause, the word “delve”, and any paragraph that could be reshuffled into a different article without anyone noticing. Also watch for the tell-tale evenness: every paragraph the same length, every section the same shape. Real writers are lumpy. They give one point 300 words and dismiss another in a line.
The fastest fix is to introduce an opinion, a specific number from your own experience, or a mild admission of doubt. “I’ve tried this four times and it worked twice” is a sentence no model will produce unprompted, and it makes everything around it feel more true.
When writing overlaps with study and work
A lot of people arrive at ChatGPT for writing because of an essay, a report or a client deliverable. That’s a different situation from a blog post — the stakes are institutional, and the rules are usually unwritten. If that’s your context, it’s worth reading about how to use ChatGPT as a student without cheating yourself out of the skills you’re paying to learn. The same logic applies to workplace writing, where the temptation to hand over first drafts wholesale tends to cost people their edge quietly — a theme that comes up in this breakdown of where ChatGPT saves business hours and where it quietly costs you.
The short version: use it to think, not to substitute for thinking. Ask it to argue against your position. Ask it what a sceptical reader would object to. That’s genuine help. Asking it to write the thing is how you end up with a document nobody, including you, believes.
It isn’t the only tool, and it isn’t always the right one
ChatGPT is a generalist. For long-form editing, the long context windows and tone controls elsewhere sometimes win. For factual research with citations, a search-native tool will save you an hour of verification. For fiction, some writers find Claude’s prose more natural and less prone to tidying everything into neat paragraphs.
If your work is mostly a single job — editing, research, drafting at volume — it’s worth comparing options rather than defaulting. This roundup of ChatGPT alternatives picked by what you actually need is organised the sensible way, by task rather than by hype.
The rule I keep coming back to
Use ChatGPT for the parts of writing that feel like admin, and do the parts that feel like thinking yourself. Outlines, variants, cuts, titles, subject lines, formatting — let it grind. Argument, evidence, structure, the weird specific detail that only you noticed — keep those.
The writers who get real value out of this tool aren’t the ones with the biggest prompt libraries. They’re the ones who know exactly which 20% of the job to hand over and which 80% to protect. That ratio takes a few weeks to feel out, and it’s different for everyone. But once you find yours, the blank page stops being the enemy — and the writing still sounds like you, which is the only part that actually matters.