The art of talking to it

Here we are. You know that the machine does not think, that it arranges words by address, guesses what comes next, has been trained to please, sometimes invents, and wears spectacles.

The bonnet is closed. Place your hands on the keyboard.

One question remains, and it paralyses a great many people. How exactly do you talk to it? Do you need keywords, as you would with a search engine? Special formulas? Should you say hello?

In the field, the question you type has a name: the prompt. There is no need to look for another word; everyone uses the English term. You will also hear about prompt engineering, presented in some quarters as an arcane discipline reserved for initiates, whose secrets can be yours after three days and a few hundred pounds.

Let us be frank.

Talking to a machine is not a technical skill. It is a communication skill, and you already possess it.

Forget computing. To get something worthwhile from your assistant, think of it as a very unusual trainee. Brilliant, capable of reading a thousand books a second, equipped with an immense vocabulary. And suffering from total amnesia where you are concerned.

It does not know who you are. It knows nothing about your profession, your habits, your audience, what you have already tried or why the task matters. It knows nothing you do not write down.

Tell it, “write an email cancelling the meeting”, and it will write an email. Curt or warm? Addressed to your director or your brother? Three lines or a full page? It has no idea, so it does what it knows how to do: it guesses the most probable continuation. You receive an average, lukewarm and flavourless answer. And you conclude that the tool is overrated.

The problem is not the tool. You gave it a quarter of the information.

Turning a bland answer into a useful one simply means filling those gaps. Four pillars are enough.

Role · Context · Format · Precision

The role. The machine loves dressing up. Tell it to act as a lawyer and it will draw from the legal neighbourhood of its map. Tell it to act as a nursery-school teacher and it will choose simple words and a reassuring tone. You are not giving it an identity. You are pointing it towards the region where it should look for words.

The context. Tell it about the situation. Explain why you are doing this, for whom, under what constraints and what you have already tried. This is by far the pillar that offers the greatest return, and the one everyone neglects. A request without context produces a generic text, because generic is precisely what is most probable.

The format. A table, a list or a paragraph? Three lines or two pages? It cannot guess your visual expectations, and without instructions it will do whatever suits it, which means headings and bullet points everywhere.

The precision. Say what you want, and also what you do not want. No jargon, no superlatives, no conclusion that summarises everything above it. Prohibitions work very well, provided they are explicit.

Let us see what this looks like in practice.

You need to tell your tenants that the rent is going up. A delicate task. Most people will write: “Draft a letter telling my tenants that their rent is increasing.” The predictable result is a cold, slightly threatening administrative letter that you will never send.

Now try it with the four pillars:

Act as a considerate but firm landlord. I need to tell my tenants, a young couple with whom everything has gone well for three years, that their rent will increase by £30 next month because the building’s service charges have risen. Write a short, warm message. Keep the tone reassuring, avoid legal language, and invite them to call me if they have any questions.

The difference is striking. You have gone from spectator to conductor.

Here is a second, more everyday example. You have some courgettes, eggs and Parmesan left. The beginner’s prompt, “give me a recipe using courgettes, eggs and cheese”, will produce the most ordinary bake on the internet, because statistically it is the most probable.

Act as an Italian chef. I have courgettes, eggs and Parmesan. I have twenty minutes before the children come home, and they hate overcooked vegetables. Give me a simple recipe in very short steps. Use nothing else except olive oil, salt and pepper.

In four sentences, you have eliminated thousands of recipes that take too long and directed the machine towards a precise region of what it has read.

Prompt

What you write in the chat window. Nothing more, nothing less.

Few-shot

Giving two or three examples of the result you expect. It is the most effective technique of all, and requires no special skill: showing is better than explaining.

System prompt

The permanent instructions you set once in your assistant’s preferences, which then apply to all your conversations.

It is time to break the trainee metaphor, because it has an important limitation.

A real trainee learns. Explain on Monday that you hate exclamation marks, and on Tuesday they remember. This one does not. Recall Chapter 4: its settings are fixed, and it learns nothing from you. Everything you explain lasts only for the current conversation, within its context window.

This has a very practical consequence. Do not start a new conversation for every question. A substantial discussion in which you have already provided context, corrected the tone twice and shown an example produces far better results than an isolated request, because all that work is still before the machine’s eyes. And if your preferences are permanent, write them once in your assistant’s settings. That is precisely what the system prompt is for.

Now for the most important point in this chapter, and the one people always forget.

The four pillars improve your first request. But the result that matters never comes from the first request. It comes from the third, the fifth or the tenth.

Treat the first answer for what it is: a draft. Do not approve it; work on it. Too long, cut it in half. The second paragraph sounds wrong, rewrite it. That word does not sound like me, find another. Give me three versions: one restrained, one funny, one very direct. You invented that statistic, remove it.

This is where everything happens, and also where Chapter 1 finally proves useful. There is nobody on the other side. Nobody whose feelings need sparing, nobody to offend, nobody who will judge you for asking for the same thing four times. You can say, “no, this is poor, start again” without the slightest guilt, and it is often the most productive sentence in the conversation.

One final tip, worth every prompting guide in the world. When you cannot quite formulate your request, reverse the roles: ask the machine to question you before answering. “Before writing anything, ask me the five questions you need answered.” You will discover that half the work was simply knowing what you wanted.

And whatever happens, keep the reflex you learnt in Chapter 5. The more precise the answer, the more polished it looks, the more figures, names and references it contains, the more carefully you must check it. Perfect form remains its most elegant trap.

You may be wondering whether all this applies to every assistant. Are ChatGPT, Claude, Gemini, Le Chat and the others really the same?

Under the bonnet, yes. The same principle, the same map, the same guessed next word, the same hallucinations, the same biases. What changes is the material they have consumed, the training they received, their house rules and the tools connected around them. Hence their rather different personalities: one more literary, another more academic, one more cautious, another more adept at code. Everything you have just learnt applies to them all, and the best way to choose remains trying two of them for a week with the same requests.

You have reached the end of the guide.

You have not learnt how to design an artificial intelligence, and that was never the goal. You have learnt what it does when it speaks to you, and why it behaves the way it does. From now on, when it flatters you, you will know that training is responsible. When it invents a reference, you will know that it is merely finding the next word. When it serves up an executive in a grey suit, you will know where the image came from.

You are no longer impressed. That is exactly what I wanted.

Remember the metaphors I have given you, and above all remember where each one breaks down: the parrot that does not copy, the map that moves, the mirror that distorts, the trainee who learns nothing. They have carried you this far, but they are not reality. The day they are no longer enough, you will be ready for what comes next, and it is waiting for you in books.

For now, you have something better than a five-hundred-pound course. You understand. Open a window, type something foolish, correct it, start again.

Nobody is watching.