A weak AI draft tells me where the problem appeared. It does not tell me whether some of your information was lost, the direction was unclear, or AI was left to make a decision that still needed to come from you.
When a draft feels off, the obvious place to start is the draft. You edit the sentence, jump back to Chat to explain the angle again or scroll to the top to check what you’d said in your prompt because that is the part you can see.
But seeing where the problem appeared is not the same as knowing what caused it.
I learnt that years ago when I was getting to grips with complex spreadsheet formulas. I had never used them before, so when something went wrong, I didn’t have enough experience to look at it and know why.
But my now-husband, who was sitting a few feet away from me, did.
I can still remember looking at him from the corner of my eye, giving a big outward sigh every few minutes and thinking, “Come on... you know the solution. You could just tell me.” He could clearly see my frustration; it was a daily occurrence at this point. I was eager to complete the task, so I just thought it would have been quicker if he had simply pointed out what I had done wrong.
But he didn’t.
And we’ve talked about his approach multiple times, because he didn’t leave me to struggle because he was being unkind. He had learnt new skills in working environments in a similar way and knew the value of working a problem out for yourself. He wasn’t feeding me clues or guiding me through questions. He gave me the space to sit in it.
So I sat in the problem.
I looked at what I had done and tried to work out what could be causing it. Had I missed something? Was one part affecting another? Could it be this, or that? I knew he wasn’t going to hand me the answer, so I had to assess the possibilities myself.
When I eventually found the cause and fixed it, I was elated. Not just relieved that the formula worked, but pleased that I had identified what was wrong.
It wasn’t simply a lesson in being patient. I learnt not to treat a problem as one solid thing. Something had failed, but there were several places it could have started, and I had to work out which one I was dealing with. That is still how I approach problems now, including the work I do around AI content.
When I look at a weak draft, I don’t immediately decide that the prompt is bad. It may be the problem, and it is a sensible first place to check, but it is only one possibility.
A better prompt cannot settle an angle that hasn’t been decided, supply information that was never included or take back a decision that should still belong to the person writing.
I start with what is wrong in the draft, then trace that part backwards. A wrong angle makes me look at where it was decided. When an example is missing, I want to know whether it was ever present. If the writing sounds unlike the person, I check what real language and preferences AI had to work with.
Those questions are not a fixed checklist I apply in exactly the same way every time. They are how I separate the possibilities instead of treating the entire draft as the problem.
I recently did this while looking at the content process of someone who runs an established service business. She had ideas, a content plan and a natural way of talking through what she wanted to say. She had used AI for around a year, reviewed what came back and could get to a LinkedIn post she considered usable.
That told me something straight away. The problem was not that she had no ideas or needed me to replace the way she naturally thought. There were useful parts of her process that needed protecting.
At the same time, AI was sometimes misunderstanding the angle. She was correcting the same writing habits and changing the formatting because it kept returning separate, skimmable lines when she preferred paragraphs. Some of the context AI needed was also being left to what ChatGPT remembered from previous conversations.
It would be fair to say that her process worked because she got something usable in the end. But getting something usable is not the full measure of a good process when the same meaning has to be recovered every time.
The repeated corrections were clues, not the diagnosis.
Once I looked at the process as a whole, the likely problem appeared to sit around how the context and writing guidance were carried into the task, along with the instructions shaping the move from her spoken idea to the draft. I didn’t need to remove the useful human parts or build something enormous. I needed to form a sensible first diagnosis and create a focused test around it.
She is testing that first process now. I don’t yet know which parts will hold up in real use or what we will need to refine, and I’m not going to pretend otherwise. The test will show us whether the diagnosis was right.
You might think it would still be quicker to edit the draft, and sometimes it is. Human editing is part of writing. I’m not trying to remove it. The difference is whether you are making a thoughtful edit or repeatedly repairing something the process lost earlier.
This is also why I’m not interested in forcing people into a rigid system. I built a rigid content process for myself once, and I didn’t like the results. Diagnosis should show us what actually needs changing, rather than convincing us that everything needs rebuilding.
What I enjoy is listening to how someone already thinks and creates, working out what is useful, then looking carefully at where their information, direction or decisions may be getting lost. From there, we build and test a way of working around what is actually causing the difficulty, then refine it from what happens in real use.
The next time an AI draft feels wrong, you may not be able to see the whole cause from the draft alone. That is the point.
The visible problem is a clue, not the diagnosis.
Sometimes the draft is the problem. Sometimes it is only the first clue.
I’d love to know what questions this opens up for you. Press reply and let’s chat.
