At the start, ChatGPT wrote like the worst freelancer I had.
That isn't a complaint about the model, it's a description of the problem. I was editing freelance writers all day and the output I got from a prompt had exactly the same faults as the output I got from people. On topic, generic, missing half of what I'd asked for, and needing a rewrite before it went anywhere near a live page.
So I did the same thing I did with the freelancers. I wrote better briefs.
It took about three months. By the end I had a document of prompts that everyone on my team was using, and roughly 90% of what came back was usable as written. That figure is my own estimate of my own workflow, not something I measured, and I should say so up front.
The Job
I was one of four editors covering five affiliate sites across the UK, global and New Zealand markets. Casino reviews, bonus pages, listing pages, blogs and guides. High volume, small team.
Two things made this harder than ordinary content work. UK gambling content sits under regulatory scrutiny, so there is language you simply cannot use. And it is YMYL content, which means Google holds it to its strictest quality standards, so a page that reads well but says something wrong is a real problem rather than a small one.
The freelance side is the relevant part. On my projects we generally got the weaker writers, so I was rewriting heavily before anything went live. That turns out to matter, because it is where the whole method came from.
What It Looked Like at the Start
My first prompts were short. Something like "write a bonus review for this casino using this information."
The output was robotic. Sometimes it didn't answer what I had actually asked. And sometimes I just didn't like it, which isn't a technical failure but still meant a rewrite.
So I would fix it by hand and move on. That went on for a while before I noticed I was fixing the same things every single time.
The Thing That Changed
The two edits I kept making were always the same. I was taking out unsafe language, words like win and life-changing that you cannot put in front of UK gambling readers. And I was putting geo-targeting back in, making sure the content spoke to British readers rather than a generic audience.
Every time. So I stopped fixing it afterwards and put it in the prompt instead. Roughly: use simple and safe language, no win, no change your life, no opportunity to win, and use geotargeting words like UK and British while varying between customers, players, punters and users.
That one instruction removed most of my editing workload. Not because it was clever, but because it was specific about something I already knew I wanted.
That is the method, and to be fair it's the only real insight I have from any of this. Whatever you keep correcting by hand is what belongs in the prompt. Your edit log is your spec. You just have to pay attention to what you're actually doing.
Where It Stopped Being an Executor
Later I added a line asking that the content be helpful, contain precise numbers and values, and offer a unique insight to the reader.
I put it in for quality control. What happened was different. The model started producing angles I hadn't thought of. Not constantly, but often enough that I would read the output and carry something from it into the next brief.
That was the point where it stopped being a thing that wrote down my ideas and became something I was working alongside. I don't want to oversell that. It wasn't having original thoughts. But asking for insight rather than only accuracy changed what came back.
The Dead End
I spent a while trying to make the prompts shorter.
Everything I was reading suggested there was some clever short instruction that would do the work of a long one. There wasn't, at least not for what I needed. Every time I trimmed a prompt the output got worse and I ended up doing the editing by hand again.
What worked was the opposite. Long prompts, as much detail as I could fit in, and time spent beforehand working out what I actually wanted before writing a word of it.
That last part is the bit people skip. Most of the work in a good prompt happens before you open the chat window. You have to know precisely what you want, which is the same problem as writing a brief for a person. If you can't explain it to a freelancer, you can't explain it to a model either.
Why I Never Had a Hallucination Problem
This is the part people tend not to believe.
Across that whole period I never had the model invent a fact that made it onto a published page. Not because it's reliable, but because I never asked it to know anything.
Look at how the prompts were built. Every one of them either handed the model the information or pointed it at a live page. Based on this information about the casino's bonuses. These are the new casinos and the details of their offers. Crawl this page and complete the structure.
Wagering requirements, payout times, game counts, provider numbers, all of it came from me or from the source page. The model was arranging and phrasing facts that had already been verified. It was never the source of truth for anything.
For the cases where I had forgotten to supply something, I added an instruction telling it not to complete missing values itself, but to leave a note for me in bold text instead.
That one was written for the model so that I could see at a glance where my own input was still missing. It turned gaps into something visible rather than something invented.
None of this is sophisticated. It is just deciding that the model does structure and language, and a person does facts.
Making It Criticise Itself
The other thing worth mentioning is that I stopped using it only to write and started using it to review.
I would have it score every heading on a page from 0 to 10 on relevance to search intent, comprehensiveness, originality, trustworthiness and user experience. Then separately on overall quality against what competitors were publishing.
The first pass was always too generous. Models want to agree with you. So I would send it back and tell it plainly that it wasn't being critical enough, and ask it to analyse the same headings again in a more critical manner.
The second assessment was usually the useful one. Knowing that the first answer is soft, and then asking again harder, got me more than any single writing prompt did.
What It Never Solved
There was still about 10% I had to do myself, and it was the same three things every time.
Repetition. It reuses phrasing across sections in a way you only notice when you read the full page in one go.
Rigidity. The output was often correct and stiff at the same time. Technically fine, not something a person would enjoy reading.
And humanising, which is really the previous two plus something harder to describe. I had a separate prompt for it, asking it to condense and humanise while keeping the language simple and user-oriented, and it helped, but it never got all the way there.
I don't think that 10% is going anywhere. It's the part where you decide whether something is actually any good, and I haven't found a way to hand that over.
What Actually Changed About the Job
The loop I was running with the model was the same loop I was running with freelancers. Write the brief. Read the output against the brief. Send it back or fix it yourself. Note what went wrong. Improve the brief.
The difference is the speed of the feedback. A freelancer takes days and you get one revision out of it. The model takes seconds and you can go round twenty times in an afternoon.
So the work didn't disappear, it moved. I spent less time writing and more time deciding what good looked like and how to specify it. Which, in general, is the more useful skill anyway.
The Short Version
- Keep a close watch on what you keep fixing by hand, then add it in the prompt.
- Don't be lazy with your prompts, be as thorough as possible for the best results. Think of all the information you'd like to have before starting a task.
- Even when you think you're close to the perfect outcome, take the output with a grain of salt.
- Make AI be its own critic, then after the reflection tell it to be harder on itself.
- Be prepared to accept that perfection isn't even in the question, the last 10% will always be yours.
Written by Vlad Craciun, SEO content writer and editor. More work on the archive page, or get in touch.