AI Does What You Ask — Not What You Mean
Neil’s AI Tip #3 — AI does what you ask. Not what you mean.
I gave a demo at a Howbeck team meeting where I asked ChatGPT to “improve a photo of a new member of staff and make it look like a corporate headshot.”
It did. The person came back slimmer. Skin smoother. Hair subtly different. Face very gently reshaped. Younger. Recognisably the same person, but not really them.
That’s what AI thinks “corporate” looks like. Trained on millions of polished stock images, it gives you polished stock. Which is the opposite of what most people want on an About Us page — patients want to see the actual person just looking their best, not an idealised avatar.
I tried the same job with a different prompt. Long, structured, deliberately constrained. It started with what I wanted. Then a list of things it must not do — don’t change facial features, don’t slim or reshape, don’t alter hair or clothing, don’t smooth wrinkles, don’t apply a beauty filter. Then what it was allowed to do — adjust lighting, correct colour, reduce noise, replace the background with a plain neutral one. And a final instruction: if any change might risk altering the person’s identity, don’t make it.
Same photo. Same tool. Completely different result. The person looked like themselves on a great day with a photo taken by a professional photographer with proper lighting not a quick snap on a phone.
The lesson isn’t really about photos. It’s about how AI handles ambiguity. If you leave space for the model to decide what you mean, it will — and its idea of “corporate” or “professional” or “tidy” will reflect the average of what it’s seen, which is rarely what you want.
The fix is specificity. And the part most people miss is being specific about what you don’t want. “Make this email more professional” gives you something corporate-bland. “Make this email more professional, but keep it warm, don’t make it formal, and skip the standard sign-off clichés” gives you something you can actually send.
A couple of related habits I’ve picked up.
If I don’t know how to write a good prompt, I ask Claude to write one for me. “I want to do X — help me draft a prompt that gets the best result, and ask me anything you need to know first.” It turns out AI is reasonably good at writing prompts for AI. The conversation that follows often surfaces things I hadn’t thought about.
And before letting the AI run with anything substantial, I ask it to play back what it thinks I want. “Before you write the report, tell me what you understand the task to be, what assumptions you’re making, and what you’re planning to leave out.” If it’s heading in the wrong direction, I’d rather find out in two sentences than after 2,000 words.
All three are versions of the same insight. The model will fill in gaps. If you don’t want it filling them, close them yourself — or ask the model to help you find them first.