The first time I asked an AI to edit a piece of my writing, I told it: "Don't make it sound too formal, don't strip out my voice, and don't make it longer."
What came back was stilted, cautious, and somehow both too formal and too flat. I'd given it three negatives and no positive direction, and it had hedged its way into exactly the kind of grey mush I was trying to avoid.
Here's the tip, plain:
Tell the AI what you want, not what you don't want.
Instead of "don't make it too formal," try "write this the way a smart friend explains things at dinner." Instead of "don't make it longer," try "keep every sentence pulling its weight." You're aiming at a target, not fencing off a minefield.
Why does this matter so much? Here's the picture I find useful.
Imagine you're giving directions to a new driver. You can say "don't take the motorway, don't go through the town centre, and don't turn left at the crossroads." Or you can say "take the B-road north, past the farm, and park by the church." Both sets of instructions theoretically describe the same destination. But the second one gives the driver an actual route. The first one just produces anxiety and wrong turns.
AI works on something close to the same principle. It doesn't experience anxiety, obviously — but it is fundamentally a prediction engine, always reaching for the most probable next move given the context you've set. When you load that context with negatives, you're telling it what space to avoid but leaving enormous ambiguity about where to land. The output becomes a kind of nervous meandering: technically not-wrong by your stated criteria, but not right either.
When you describe what you do want — a tone, a shape, a specific effect — you narrow the space of likely responses in a much more useful direction. You're not just removing bad options; you're actively pointing at good ones.
There's a behavioural science angle here worth naming. It connects to something called present bias — our tendency to feel immediate aversion to loss far more sharply than we feel the pull of a future gain. When we draft prompts, we often write them emotionally: "don't ruin the thing I've already made." We're protecting what exists rather than building toward what we want. That instinct is understandable. It's also a terrible prompting strategy.
The fix isn't to suppress the instinct. It's to translate it. Every "don't" in your prompt is a fear in disguise. Ask yourself what you're afraid of, then turn that fear into a positive description. Afraid it'll sound too corporate? Say "write this like a person, not a policy document." Afraid it'll lose nuance? Say "preserve the qualification in the third sentence — it's doing real work." The underlying concern stays. The phrasing changes from avoidance to direction.
I'll be honest about the complication here: this takes a few seconds more. Writing a negative constraint is easy — it requires almost no imagination, just a reaction to what you're already worried about. Writing a positive one requires you to have, at least partially, a vision of what good looks like. Sometimes you don't yet have that. Sometimes you genuinely don't know what tone you want until you see three examples, so you say "give me three different versions" and let the options teach you your own preference. That's a valid strategy. But "I'm not sure what I want" and "tell it what you want" aren't incompatible — "give me three options that take different risks" is still a positive instruction.
The deeper truth is that precision in prompting is mostly about precision in thinking. AI has a way of surfacing the vagueness we haven't noticed in our own intentions. When the output is wrong in a weird way, it's often because our instruction was wrong in a weird way. The prompt is a mirror.
I started by describing how my first instinct was to fence off everything I was afraid of and hope the AI would find the meadow in between. It didn't. Once I started pointing at the meadow directly, the outputs got dramatically better — not because I became a prompt wizard, but because I stopped outsourcing the thinking and started doing half a sentence's worth of it first.
That half-sentence of intention is the whole skill.
If you want to build that instinct properly — not just pick up one tip and forget it, but actually change how you work with AI day to day — that's what The AI Habit is for. It's 15 minutes a day, and the first 14 days are free. Small practice, compounding fast.
https://harrysharman.com/projects/ai-habit/