Most people who are disappointed in ChatGPT or Claude don't have the wrong model. They give the wrong instruction. They type "write a quote" and get a grey, generic text back, and conclude AI is overrated. But the model did exactly what was asked: something generic, because nothing specific was said.
The instruction you give an AI model is called a prompt. And the difference between a mediocre and a usable result almost always sits in that prompt, not in the model. The good news: prompting is a skill, not a talent. You can learn it in an afternoon.
What is a prompt?
A prompt is the text with which you tell an AI model what you want. That can be a question ("what is the difference between revenue and profit"), an instruction ("summarise this email in three sentences"), or a complete working brief with background, examples, and the desired format.
The word means something like to prod or to cue. That captures it: a prompt is the starting signal that determines what the model does. An AI model has no idea what is in your head. It only has the words you type. The more relevant information they contain, the closer the answer gets to what you mean.
Many people treat a prompt like a Google search: short, a few keywords, then see what comes out. That is exactly the wrong reflex. Why that is, and how to hold a conversation with AI instead, I wrote in why you should be patient with AI. This guide is about structure: what a good prompt looks like.
The anatomy of a good prompt
A strong prompt almost always contains the same building blocks. Not every prompt needs all of them, but the more important the task, the more you include.
Context. The background. Who is this for, in what situation, what has already happened. "I run an installation company with eight engineers" gives the model a very different starting point than nothing at all.
Role. Who should the AI be? "Act as an experienced copywriter" or "you are a bookkeeper explaining something to me" steers tone and level of detail surprisingly strongly.
Task. What exactly should happen, in verbs. Not "something about marketing" but "write three subject lines for an email to existing customers".
Format. What should the answer look like? A table, a list, five bullets, a text of no more than a hundred words. Say nothing and the model chooses for you, usually not what you wanted.
Examples. One example of what good looks like often does more than three paragraphs of explanation. Paste a previous quote you liked and say "in this style".
Two frameworks that give you something to hold on to
If you would rather not think about those building blocks every time, use a fixed template. Two that work in practice:
CRAFT. Context, Role, Action, Format, Test. You describe the situation, give the AI a role, say what it must do, in what form, and which criteria the result must meet. Handy for one-off tasks.
CRISP. Context, Role, Instructions, Specifications, Persona, especially useful for recurring business tasks where you want the output to feel the same every time. You fix who the AI is and how it should sound, so you don't have to type that again each time.
Which template you pick matters less than the fact that you use one. It forces you to give the information the model needs, instead of hoping it reads your mind.
Five examples: before and after
Nothing shows the difference better than the same task, twice.
Answering a customer email. Before: "Write a reply to this complaint." After: "A customer complains that their order is three days late. Write a reply of no more than a hundred words, friendly but without over-apologising, offer a concrete solution, and end with a question asking whether that works for them. Write in informal English."
A job ad. Before: "Make a job ad for an engineer." After: "Write a job ad for a maintenance engineer at an installation company. Audience: vocationally trained, 25 to 40, values a stable team and short lines over a high salary. No more than three hundred words, no clichés like 'dynamic environment', end with how to apply."
Explaining figures. Before: "Explain these figures." After: "Here is my monthly revenue over a year. Summarise in three sentences what stands out, name the strongest and weakest month, and give one possible explanation I could look into myself. Don't present assumptions as fact."
Developing a blog idea. Before: "Write a blog about AI." After: "Write an outline for an 800-word blog for SME owners who don't use AI yet. Topic: where they are best off starting. Practical, no hype, with one concrete first step. Give only the headings and one sentence per heading."
Translating while keeping tone. Before: "Translate this into English." After: "Translate this product description into British English. Keep the informal, enthusiastic tone, but avoid literal translations of Dutch expressions. Audience: British webshop customers."
The pattern is always the same: you replace "do something" with "do this, for this person, in this form, and watch out for that".
The mistakes almost everyone makes
Too short. By far the most common. A five-word prompt gives a five-word-quality answer.
Wanting everything at once. A prompt that asks for a strategy, a plan, and the copy usually delivers all three half-baked. Break it into steps.
Not saying what you don't want. "No jargon", "no longer than a hundred words", "no invented figures": the negative instructions are often as important as the positive ones.
Not following up. The first answer is a draft, not a finished product. "Make it shorter", "write it more formally", "give three variants": that is part of it.
Giving no example when you have one. If you know what good looks like, show it.
How do you save and reuse good prompts?
A prompt that works well once is worth gold, as long as you can find it again. A lot of time saving is lost because people start from scratch every time. So keep a simple file or note with your best prompts per recurring task: the quote prompt, the email prompt, the summary prompt. Tools like ChatGPT and Claude also have built-in features for this, such as fixed instructions or saved projects, so you don't have to retype your context each time. Which tool suits that best depends on your work; I put the options side by side in ChatGPT alternatives for SMEs.
2026: from prompt engineering to context engineering
An important shift to know. Models have become a lot better at understanding what you mean over the past few years, even with imperfect phrasing. At the same time it has turned out that the biggest gain sits not in the exact wording of a single question, but in everything the model knows at the moment you ask it.
That is called context engineering. Instead of polishing the perfect sentence, you make sure the AI has the right background to hand: your previous documents, your house style, examples of good work, the conventions in your company. For SMEs that means, in practice: the more relevant material you provide, the less it matters whether your prompt is grammatically perfect. The question shifts from "how do I phrase this cleverly" to "what does the AI need to know to help me well".
How do you get better at this?
By doing it. Take a task you had to do this week anyway, write a prompt for it using the building blocks above, and improve the result over two or three rounds instead of leaving it at the first attempt. After a week or two it feels natural.
If you want your team to master this on your own work, with examples from your own practice, that is what an AI training is for. And for the mindset behind good prompting, the patience to hold a conversation instead of firing off a single instruction, read why you should be patient with AI. And if you want to know first where AI can save you the most time, the free AI scan gives a first analysis based on your own situation.
A good prompt is not a trick. It is simply writing down clearly what you want, for someone who can't read your mind. That is a skill that comes in handy outside AI too.
Curious what AI can do for your business?
Take the free AI Scan and find out in 1 minute.