A good prompt has four layers: you give the AI a role, you provide context, you spell out the task precisely, and you state the expected format. Most unusable answers aren't caused by a weak model — they're caused by one or more of these layers being missing. When that happens, the AI fills in the gap itself and produces generalities.
This article walks through the four layers with concrete examples and shows what to avoid.
What is a prompt? A prompt is the instruction you give the AI. It isn't a question and it isn't a magic word — it's a short work order in which you state whose perspective to take, what background to work from, what exactly you're asking for, and in what form. The less the AI has to guess, the more precise the answer.
Why does the AI give a generic answer?
Because it received a generic question. If you type "write me a proposal," the AI has no idea who you are, who you're writing to, what you're selling, or what tone you normally use — so it returns the most common, most neutral pattern it knows from its training data.
The same model, the same subscription, gives a noticeably better answer once all four layers are in place. The difference isn't the tool — it's how much information you gave it to work with.
1. Role: whose head should it think with?
The role sets the perspective and the professional level. It's not decoration — it meaningfully narrows the vocabulary and the frame of reference the model works from.
Bad: "Help me write a letter." Good: "You are an experienced proposal writer at a construction contractor."
A good test for a role: could you use it to describe a real person in a job posting? If yes, it's specific enough.
2. Context: what does it know about the situation?
Context is what you know that the AI doesn't: who the client is, what's happened so far, what's at stake, what the constraints are. This layer is what separates boilerplate text from something that's actually about you.
Good context, for example: "The client has 40 employees, we've worked with them twice before, and payment was late both times. Now they're asking for a quote on a bigger job."
With this, the AI already knows the deposit question can't be left out — without it, the AI would never have raised the issue.
3. Task: what exactly should it do?
The task should be a single, unambiguous verb. If a prompt asks for four things at once, all four typically come out mediocre.
Bad: "Review it, fix it, shorten it, and add a subject line." Good: "Write a cover letter for the attached quote line items."
If several steps are needed, split them into separate prompts and build on the previous answer. This multi-step approach is one of the biggest jumps between beginner and experienced use.
4. Format: in what form do you want it?
Format saves the most rework afterward. Length, tone, structure, restrictions — it's all covered here.
A good format constraint: "Maximum 150 words, informal tone, ending with one question. Don't use a bulleted list."
If you don't specify a format, the AI falls back on its own default: long, bulleted, neutral text. That's rarely what you actually need.
What makes a prompt reusable?
Turning it into a template. Once a prompt works, replace the variable parts with placeholders — client name, amount, deadline — and save it. Next time, you just fill it in.
This is the point where prompting turns from an individual skill into a company asset: your colleagues can use the templates that work without having to walk the whole learning curve themselves. In the prompt engineering masterclass this is exactly the step we practice through — by the end of the day you'll have your own prompt library, written for your own tasks.
The four most common mistakes
Overly polite, overly long preamble. The AI doesn't need an introduction and doesn't need to be asked nicely. Unnecessary sentences pull attention away from what matters.
Multiple tasks crammed into one prompt. Break it into steps. Multi-step work isn't slower, it just looks different.
Not specifying what it shouldn't do. A restriction is often more effective than an instruction: "don't invent data that isn't in the text" — that's a single sentence, and it prevents a lot of wrong answers.
Accepting the first answer. A good prompt is rarely right on the first try. Ask for specific revisions: "this is too formal, make it more direct, and cut the last paragraph."
What if the whole team should use this?
Write down in a document what may and may not be given to the AI, and share the templates that work. Without this, every colleague reinvents the same wheel, and quality depends entirely on how skilled each person happens to be.
If you want to roll this out across your organization, that's what AI training is for: the team practices on its own tasks and takes home working templates, not just theory. And if you're not even sure which process to start with, AI consulting starts by assessing where the payoff would be fastest.
In summary
A good prompt isn't longer, it's more complete. Role, context, task, format — once all four are in place, the same tool delivers a noticeably better result, and the answer becomes repeatable instead of a matter of luck.