Prompt engineering

In the previous step I separated the prompt from the dialogue: each turn you write is a one-off message, and the quality of what you get back depends a lot on how you draft it. Here I stop right on that, on drafting it well. It took me a while to realise that almost all my bad results didn't come from the machine, but from how weak my requests were. I want to save you that detour.

What you'll understand by climbing this step is that asking well isn't a gift or a trick: it's structuring what you have in your head so that nothing important goes unsaid.

The name is scarier than the thing

This is called prompt engineering, and the term is imposing: it sounds like a technical discipline, like something reserved for initiates. When I first saw it I took for granted there'd be a special syntax to learn, commands, some hidden formula. There's none of that.

Prompt engineering is just drafting what you ask the AI in an orderly, unambiguous way. Nothing more. What surrounds it with mystery is the name, not the practice. And understanding this from the start takes off your back the most damaging idea circulating about the topic, which is the following.

The magic-words misunderstanding

Everywhere you'll see lists of "secret prompts" and phrases promising to unlock the AI, as if there were passwords that force it to behave better. I collected a few. I'd paste them as they were, and sometimes they seemed to work, but I never knew why, so I couldn't repeat the success either.

The underlying problem is that those lists treat the prompt like a spell. And it isn't. The AI doesn't respond to words with power, it responds to what's written. Remember what we saw in the first step: it doesn't guess your intention or understand what you leave unsaid; it works with the text you give it and fills the rest with the most likely thing. If you leave a gap, it covers it its way, not yours. Asking well isn't finding the right incantation, it's not leaving gaps.

The five pieces of a good request

When I stopped looking for formulas and noticed what my good requests had in common, the same five pieces always came up. I'll name them one by one, because having them in your head is half the battle.

The role is who you want the AI to speak as. If you tell it "act as a customer service manager," it steers the tone, the vocabulary and the judgment it's going to answer with all at once. The task is what you want exactly, said with a clear verb: draft, summarise, translate, compare. It's the piece that's never missing, and even so it's the one most people leave half-done.

The context is the background information the machine needs and that you take for granted without realising. That the customer bought ten days ago, that they're angry, that it's the second time they've complained: the AI knows none of that unless you tell it. The format is how you want the answer: in what tone, at what length, as a list or in prose. And the examples are a sample of what you expect; a single sample steers more than three sentences of instruction, because you show it the result instead of describing it.

The same task, before and after

I see it more clearly with a case. Imagine you write this: "write me an email." It's a task, yes, but it's almost empty. The AI doesn't know what the email is about, or to whom, or in what tone, so it fills all those gaps on its own and gives you back something generic that's no use to you. It isn't that it gets it wrong: it's that you gave it a huge gap and it covered it as best it could.

Now the same request with the pieces in place: "Act as a customer service manager (role). Write an email apologising for a delay in the shipment (task). The customer bought ten days ago and is upset (context). Warm tone, eighty words maximum (format)." It's the same task, but now there's almost nothing left for the machine to guess. The jump in quality between one version and the other is enormous, and I haven't used a single magic word: I've only said out loud what I used to keep to myself.

Why this really works

It helps to understand the mechanism, not just the recipe, because that way you'll know when to apply it. The AI builds its answer leaning on what you put before. The more complete your request, the less ground it has to fill in blind, and the less room there is for it to drift toward the generic.

Each piece tackles a different gap. The role fixes the style from the start. The context takes away the catalogue answers, the ones that would fit anyone. The format saves you having to rewrite or trim afterward. And an example marks the target at a glance. They aren't decorations: each one closes a door it would otherwise slip through to make something up.

A toolbox, not a form

Now a nuance, because otherwise it would seem you always have to fill in the five boxes and that's not so. For something simple, the task and the format are more than enough: "translate this paragraph into English for me, in a formal tone" doesn't need a role or examples. Forcing the five pieces into a simple request only makes it heavier, without improving the result.

Think of the five pieces as a toolbox, not a mandatory form. When an answer comes out weak, run through the list mentally and ask yourself which piece you missed: almost always it was the context you took for granted, or a format you never got around to asking for. With that you already have the frame of any good request. What comes after is learning to raise that frame effortlessly, without having to think it through from scratch each time.

Definitions

- Prompt engineering: drafting what you ask an AI in an orderly, unambiguous way. Despite the technical name, it requires no special syntax; it's just saying well what you want. - Prompt: the message you send the AI in a turn. It's the text the machine leans on to build its answer. - Role: the identity you want the AI to speak from ("act as..."). It steers its tone, its vocabulary and its judgment. - Task: the specific action you ask for, expressed with a clear verb (draft, summarise, compare). It's the piece that must never be missing. - Context: the background information the machine needs and that you usually take for granted. Without it, the answers come out generic. - Format: how you want the answer: tone, length, structure. It saves you rewriting what you get. - Example: a sample of the result you expect. It teaches the AI the target better than an instruction described in words.

Further reading

- Aprender21, Estructura de un Prompt Perfecto: Guía Definitiva (2026) — a breakdown in Spanish of role, task, context, format and constraints, with copyable examples. https://www.aprender21.com/blog/estructura-de-un-prompt-perfecto - Platzi, Estructura de prompts que eliminan respuestas genéricas — a didactic guide to the components of a prompt. https://platzi.com/cursos/prompt-engineering/como-estructurar-prompts-efectivos-con-r/ - Universidad de Almería (Biblioguías), Creación de prompts — brief, solid academic material on how to formulate requests. https://biblioguias.ual.es/IA/prompts