The prompt vs the dialogue

In the previous steps we saw what's on the other side and how it builds its answer chunk by chunk. Now begins the part that most changed my way of using the chat: how you talk to it. And before fine-tuning a single word, you have to understand that you have two ways of working with it, not one.

This step opens the stretch on talking to it well. I won't get yet into how to draft the request better; here I only want you to see the difference between firing off an order and sustaining a conversation.

Two ways of asking

The first way is the one we all use at the start. You write what you want, hit send and keep whatever comes out. If it serves, good; if not, you start again from scratch with another sentence. I call this the one-off prompt: a single shot, a message that goes and comes back with no follow-up.

The second is another thing. You read what it returned, see what's off, and instead of throwing it away and rewriting from the start, you tell the chat what to fix. "Shorter." "Drop the example at the end." "The second paragraph doesn't add up, redo it." And you carry on like that, intervention by intervention, until the answer resembles what you wanted. That's dialogue, and it almost always gives a better result than the single shot.

When I understood the second way existed, I stopped fighting with the chat. The problem wasn't me writing badly: it was that I was using the tool by halves.

The myth of the magic prompt

There's a very widespread idea worth dropping as soon as possible. Many people believe there's a perfect prompt, an exact formula that, if you nail it on the first try, gives you the good answer all at once. And, as a consequence, they think that when the answer doesn't serve it's because they wrote the order badly.

Sometimes it's true the order was weak. But most of the time it isn't. The machine's first attempt is almost always a draft, not because you got the request wrong, but because the chat doesn't yet quite know what you have in your head. You do; it doesn't. The first answer is its best guess with the little you gave it, and it almost never hits the bullseye blind.

Knowing this changes the attitude. If you expect a perfect hit on the first shot, you get frustrated when it doesn't come and start from scratch over and over, each time with a different sentence, flailing. If you take for granted that the first attempt is a draft, you read what came out and start correcting it. It's much less effort and the result is better.

Why dialogue works better

Here you see what we saw in the first two steps. Within a single conversation, the AI has in view everything said before: your first message, its answer, your correction. It doesn't start clean with each turn. When you tell it "shorter," that indication doesn't float alone, it adds to everything before and reorients what comes next.

Remember how it builds the answer: word by word, leaning on the text in front of it. Well, in a dialogue that text in front of it grows and gets more precise with each intervention of yours. Each "almost, but…" adds one more clue about what you're after, and the chat composes the next attempt with all those clues together. That's why the third or fourth answer usually comes much closer than the first: it isn't that the machine has got smarter, it's that it now has a clearer idea of what you want.

Starting from scratch, on the other hand, throws all that information in the bin. If you open a new conversation every time something doesn't please you, you force it to guess again with the same scant starting information. Dialogue preserves the path travelled; the repeated shot erases it.

"Almost, but…" is the most useful phrase

It took me a while to accept that correcting isn't a sign of having done it wrong. For a while, when the answer didn't serve me, I felt I'd failed at asking. Until I realised that the adjustment is the method, not a patch.

I'm not making this up: it's the way the popularisers themselves recommend working with these models. BBVA, reviewing usage techniques, describes work with AI as a cycle in which the model proposes and you keep adjusting it, and advises asking for "a second, refined version" instead of redoing it all from scratch. The underlying idea is to treat the chat as a co-pilot you talk with, not as a vending machine you either get the coin right with or not.

So the most useful phrase I learned to say to the chat is "almost, but…". "Almost, but make it shorter." "Almost, but the tone is too serious." Each of those phrases is a little nudge that reorients the next attempt without losing what was already good. Working like this, on back-and-forth, is what separates getting something decent out of the tool from fighting with it.

When one shot is enough

I don't want to leave you with the idea that the one-off prompt is bad and dialogue is good. It would be unfair and, besides, false. The single shot has its place.

For simple, clear tasks, where what you ask leaves little room for interpretation, the one-off prompt is more than enough. "Translate this sentence into English for me." "Give me ten names for a project." "Summarise this paragraph in one line." There you don't need to converse: you ask, they give, you carry on. And if it's something you're going to repeat many times the same way, it's worth spending a while leaving that single order polished, because you'll reuse it.

Dialogue wins when the task is fuzzy, when you yourself aren't quite clear what you want until you see a first attempt, or when you're after quality and not just any answer. It isn't that one way is right and the other wrong; it's knowing which one each task calls for. The round-and-done things, in one shot; the things that need cooking, by conversing.

What comes now

From here on, the whole stretch is about fine-tuning these two forms. If you've seen that the first shot matters for simple tasks, what comes next is learning to build that shot better, so that even the first attempt lands closer. That's the ground I step onto in the next step.

Definitions

- Prompt: the message you write to the chat, your request or instruction. It's the input the model uses to compose the answer. - One-off prompt: a single message you send and keep whatever comes out of, without continuing the conversation. It serves for simple, repeatable tasks. - Dialogue: working with the chat on back-and-forth, reading each answer and asking for corrections that add to what came before. It wins when the task is fuzzy or you're after quality. - Iterative refinement: the method of improving an answer attempt after attempt, adjusting little by little instead of starting from scratch. It's the mechanics of dialogue.

No comments yet

No comments yet. Be the first.