How to write an AI prompt that gets what you actually want
Most disappointing AI answers come from a vague question, not a weak tool. Here is how to ask so you get something useful first time.
You ask it something, the answer comes back flat and generic, and a small voice wonders whether you are just not very good at this.
You are. Nine times out of ten, the answer was rubbish because the question was. A vague prompt gets a vague reply, every time, no matter how clever the tool is.
The good news, and I see this constantly, is that asking well is a skill anyone can pick up in an afternoon.
You do not need clever prompt formulas off the internet. You need to give the AI four things it genuinely cannot guess: who it is being, what you want, what good looks like, and any facts only you know.
Tell it who to be
"Write me a sales email" gives it nothing to aim at. "You are a friendly salesperson for a small plumbing firm" gives it a voice and a stance.
Setting the role focuses everything that follows. It takes five words.
Be specific about what you want
Spell out the format, the length, the tone and the audience. The more you decide up front, the less it guesses.
Vague: "Give me some marketing ideas."
Specific: "Give me five low-cost marketing ideas for a local cafe, each one sentence, aimed at attracting weekday morning customers."
The second one is almost impossible to answer badly. That is the whole point.
Show it what good looks like
If you have an example of the thing you want, paste it in and say "in this style." One good example teaches it more than a paragraph of description, and it lifts the quality of almost any answer.
Under the bonnet, the model does not read your prompt as words. It chops it into tokens, small chunks of text roughly the size of a syllable, and predicts the next one over and over. Everything it can pay attention to at once, your instruction plus the conversation so far, lives in the context window, a fixed budget of tokens. Fill it with waffle and the useful detail gets crowded out. Two more bits explain why the tips above work. There is a difference between the system message, the standing instruction that sets who it is being and holds for the whole chat, and the user message, your individual request. Putting "you are a friendly plumber" in the role rather than burying it mid-question is literally setting that system message. And pasting an example is called few-shot prompting: a couple of worked samples shift the prediction toward your style far harder than any amount of describing it. Clear role, tight context, a worked example. That is most of prompting, demystified.
Give it the facts it cannot know
The AI does not know your prices, your deadline, your customer, your last conversation. If those matter, tell it.
Half of all bad answers are just the AI filling a gap you could have filled yourself.
Then refine, do not restart
Your first prompt rarely lands perfectly, and that is fine. Do not throw it away and begin again.
Tell it what to change. "Shorter." "More direct." "Drop the last point." "Make it warmer." You are steering a conversation, not buying a vending machine snack. The second and third nudge is usually where the gold is.
Get into the habit of asking this way and the tool stops feeling unpredictable. You are no longer hoping for a good answer. You are setting it up to give you one.
If you would rather not have to think about prompts at all, and just have AI reliably do the job, building that in is exactly what we do.
Book a quick chat →Related: How to use AI without it sounding like a robot wrote it.
Common questions
What makes a good AI prompt?
Four things: tell it who to be, be specific about the format and audience, show it an example of what good looks like, and give it any facts it cannot know, like your prices or deadline. Vague prompts get vague answers.
Why are my AI answers so generic?
Usually because the prompt was generic. The AI fills gaps with the most average option it can find, so the more you decide up front, the role, the length, the tone, the specifics, the less it guesses and the better the answer.
Should I rewrite my prompt if the answer is wrong?
No, refine it instead. Tell it what to change, shorter, warmer, drop a point, and keep going. The best results usually come on the second or third nudge, not the first attempt.