1. The five building blocks of a good prompt

A prompt is the instruction you give to ChatGPT: the text in the input field with which you determine what is going to happen. The difference between an answer you throw away and an answer you can send after two small adjustments almost never lies in the model. It lies in what you have given it.

The model always fills in every gap

This is the principle everything rests on: a language model never leaves a gap open. If you do not name a target audience, it picks one itself. If you say nothing about length, the result will be of average length. If you do not specify a tone, it falls back on a smooth, faintly American corporate style with plenty of adjectives and bullet points. That is exactly the writing style readers have now learned to recognise as AI text.

There are two ways to deal with that. You can repair afterwards what you didn't want, or you can choose in advance. The five building blocks are the five places where you take that choice out of the model's hands.

Role. You tell the model what position it is writing or thinking from. "You are an experienced employment lawyer" produces different wording, different caveats and a different structure than the same question without a role. The role mainly steers terminology and angle. It is the building block that makes the least difference, but it also only costs you half a sentence.

Task. The verb everything revolves around: summarise, rewrite, compare, come up with, assess. A vague task produces a vague answer. "Have a look at this text" is not a task. "Extract the three main objections from this text" is.

Context. Everything the model cannot know and you can: who it is intended for, what came before, what decision hangs on it, what sensitivities are at play. This is the building block that is most often missing and that produces the biggest leap.

Format. The form of the output: number of words, table or running text, number of variants, subheadings or not. A tight format also saves you reading time, because you get no more than you need.

Constraints. What absolutely must not be included. No inventing assumptions, no English loanwords, no promises about delivery times, no exclamation marks. Constraints work when they are concrete and not when they describe an atmosphere.

Context: what the model cannot possibly know

The model knows nothing about your organisation, your customer, your previous email or the agreement that fell through last week. Everything you take for granted has to appear somewhere in your prompt. A useful test: imagine a temp taking over this task from you on their first day at work. What would you tell that person before they start? That is exactly what belongs in your context. Usually it comes down to four things.

• The situation: what came before? "The customer complained two weeks ago about a late delivery and has not yet had a reply."
• The recipient: who is it for and what do they already know? "A team leader with no technical background who has to explain the decision to their team."
• The purpose behind the task: what should happen after reading? "I want them to fill in the form by Friday at the latest, not to start debating the policy."
• The raw material: what source material belongs with it? Minutes, an earlier email, figures, a section of policy text.

If you paste in a longer piece of source material, put a separator in. For example the word SOURCE TEXT above the material and TASK above your instruction. Without that separator it regularly happens that the model mistakes a sentence from your source material for an instruction, and then you get an answer to a question you never asked.

Pointing out tone instead of naming it

The word "professional" means something different to you than it does to the model, and probably something different to your colleague as well. Instructions that do work are concrete and measurable: "an average of no more than fifteen words per sentence", "use the informal 'je', not the formal 'u'", "no 'we are delighted', no 'in the context of', no exclamation marks", "conclusion first, explanation afterwards". A comparison helps too: "the tone of a colleague you know well who is in a hurry".

Five mistakes everyone makes at the start

Most disappointing answers are not down to the model, but to a handful of mistakes you can quickly eliminate once you know them.

• Too many tasks in one prompt. You ask for analysis, summarising, rewriting and translating all at once. The first goes well, the rest only half. Break a prompt like that into separate steps.
• Questions without criteria. "Is this a good plan?" gets you agreement, because the model is happy to build on whatever you suggest. "Assess this plan for feasibility within three months and name the four biggest risks" gets you substance.
• Politeness instead of instruction. "Could you perhaps see whether you could do something with this text?" contains no instruction. The model then guesses what you mean.
• Negations without an alternative. "Don't make it too long" gives no direction. "Maximum 120 words" does. Whatever you rule out, add what you want instead.
• Start over straight away. A new conversation throws away all the context you have built up. Carrying on in the same conversation almost always yields more, and that is what the next chapter is about.

Summary

• The model never leaves a gap open: whatever you do not choose, it chooses for you.
• The role steers terminology and angle; the task is the verb that everything revolves around.
• Context is what the model cannot know: the situation, the recipient, the purpose and the source material.
• Format is length and form in numbers; constraints are what must not be included.
• Mark off source material with a heading, otherwise the model will read an instruction into it.
• You steer tone with measurable instructions, not with words like "professional".
• The most common mistakes: too much at once, assessing without criteria, politeness without instruction, prohibiting without an alternative, and starting over too quickly.