I spent three months writing prompts that started with "don't". Don't make it too formal. Don't use jargon. Don't make it sound like a marketing email. Every output came back as a slightly worse version of the thing I was trying to avoid, and I couldn't figure out why.
The tip is this: tell the AI what to aim at, not what to steer clear of.
That's it. One line. But the reason it works is interesting enough that it's worth sitting with for a moment, because once you see it, you can't unsee it.
Think about a golfer who's been told there's a bunker to the left of the green. Every piece of their attention is now on the bunker. The brain, trying to be helpful, has built the most vivid possible map of exactly where not to go — which means that's precisely where the club wants to swing. Negative instructions don't create a picture of success; they create a picture of failure, rendered in high definition.
An AI works similarly. When you write "don't make it sound like a press release", you've handed the model an anchor. The clearest, most concrete reference point in that sentence is press release. The model has to process that concept at full resolution before it can try to move away from it — and moving away from a single point in every direction is not the same thing as moving toward a specific destination. The output ends up in a hazy middle ground. It doesn't sound quite like a press release, but it doesn't sound like anything in particular either.
Whereas if you say "write this in the voice of a smart friend explaining it over coffee", the model has a destination. It can aim. The output has texture.
This is what psychologists call the ironic process theory — the idea that trying to suppress a thought makes it more present, not less. It's why "don't think about a white bear" immediately fills your head with one. I'm not saying AI is thinking in any meaningful sense, but the structural problem is the same: a negation gives you the unwanted thing as the reference object, and then asks the system to work away from it without a compass.
The honest complication: sometimes you genuinely don't know what you want, only what you don't want. That's real. You've read five drafts and you know the sixth needs to feel different, but you can't yet name the quality you're after.
When that happens, I do this: I ask the AI to give me three short versions — each with a named style or register — and then I pick the one that's closest to what I was trying to describe. Not to use it wholesale, but to see the options rendered, so I can suddenly articulate what I actually want. "More like option two but tighten the second paragraph" is a vastly sharper instruction than any amount of "not too formal".
The three-options move doesn't replace the tip. It's a way of arriving at the tip when you're stuck. You're still, ultimately, looking for a target to aim at.
Once I started writing toward rather than away from, something else shifted. I became more honest about what I actually wanted from a piece of writing or a piece of thinking — because the AI was forcing me to name it. Negative prompts are a form of vagueness dressed up as precision. Positive prompts require you to know your mind.
That's the real value. Not that the outputs get a bit better, though they do. It's that the discipline of writing a good positive prompt is itself a kind of thinking, a way of clarifying what success actually looks like before you've started. The AI becomes a mirror for your own fuzzy intentions.
The golfer analogy holds all the way through. The bunker isn't irrelevant — knowing it's there is useful information. But the last thing in your head before you swing should be the pin, not the sand.
If you want to build prompting habits like this one properly — not as a collection of disconnected tips, but as a daily practice that actually sticks — The AI Habit is the programme I built for exactly that. It's 15 minutes a day, the first 14 days are free, and it's designed for busy professionals who want to use AI well rather than just use it more.
https://harrysharman.com/projects/ai-habit/