How to generate very good prompts with ease

In this step I want to take a fear off you: to write good prompts you don't need to master any theory. There are shortcuts anyone can use from day one, and the best of all is to ask the AI itself for help.

In the previous step we broke down the pieces that make up a good request, and maybe you were left with the feeling that now you have to memorise them and apply them carefully every time. No. That was to understand what's underneath. For day to day, what really serves you is four or five tricks you copy and use, and that's it.

The talent that doesn't exist

For a while I thought writing good prompts was a matter of people with a knack, a gift some had and others didn't. I'd see someone get a rounded answer and assume they knew some secret formula, some magic word I was unaware of.

There's no such thing. Writing good prompts isn't a talent or a matter of memorising formulas: it's a simple craft made of repeatable tricks. Whoever seems to have the knack has, in reality, only internalised two or three habits you're about to have in front of you. And the first of them, the one that surprised me most, is that you can hand the work to the tool itself.

The trick I use most

Here's the star shortcut, the one I use almost always. Write whatever you want however it comes out, without polishing it, and add at the end a sentence along these lines: "before answering, improve this prompt and tell me what you need to know to do it well."

What happens then is small but very useful. Instead of rushing to answer your half-baked request, the AI gives you back a rewritten, clearer version of what you asked, and along the way asks you the questions it needs to get it right: "who's this for?", "what length do you want?", "do you prefer a warm or formal tone?". You answer those questions and, now yes, it delivers a result up to standard. This technique has a name, meta-prompting: using the model itself to fine-tune the instructions you give it.

It works for a reason that connects with the first step. The AI knows its own workings far better than you do; it knows what information helps it answer and what gaps leave it in the dark. By asking it to polish your request, you set it to do exactly what it's good at, and as a bonus you learn to ask better by watching how it rewrites yours.

A template with blanks to fill in

The second shortcut is to have a fixed template on hand and just fill it in. One that works very well is this:

"Act as [role]. Your task is [objective], for [audience], in a [style] tone and [format] format."

You fill in the brackets and that's it. "Act as a science-communication editor. Your task is to summarise this text about neural networks, for someone with no technical background, in a warm tone and a five-sentence format." There's nothing to invent each time: the structure is already there, you only change the filling. In time you'll know the template by heart and write it in one go, but at the start having it written in a note saves you half the stumbles.

This habit of fixing the role, the task, the audience, the tone and the format shows up again and again in the prompting guides, and not by whim: they're the blanks the AI needs to see filled so it doesn't improvise.

Show it instead of describing it

The third trick is one of the most rewarding and almost nobody uses at the start: instead of describing what you want, give it an example of it. Words to explain a style or a format always fall short; one sample is worth all of them.

If you want it to classify reviews as positive or negative, don't dump a paragraph on it explaining what you mean by each. Paste it two or three already-classified reviews —"this one's positive, this one's negative"— and ask it to continue. The AI catches the pattern on the fly and applies it to the rest. The same if you're after a specific tone: paste it a text that sounds the way you want and tell it "write me something like this." Showing always beats describing.

And when the task is large or messy, add the fourth shortcut: ask for the answer in steps. A phrase as simple as "reason it out step by step before giving me the answer" changes the result on problems that require some thinking, because you force it to lay out the reasoning instead of jumping to the conclusion. In the first step we saw that the AI builds the answer chunk by chunk; asking for steps is giving it more room to build well before committing to an output. One caveat: this pays off mainly with fast models. Reasoning models, which we'll get to later, already do that work inside, and asking them for it is redundant.

Save what works for you

There's a habit left that isn't a writing trick but one of order, and for that very reason gets forgotten. When a prompt gives you a rounded result, save it. In a note, in a document, wherever.

The point is that good prompts get reused with minimal changes. The one that summarised an article well for you will summarise the next one well by changing only the text; the one that gave you a perfect tone will serve you for another similar task. Bit by bit you gather a small collection of requests you know work, and you stop starting from scratch each time. That folder is worth more than any magic formula, because it's made to your measure and with things you've already seen work.

From the blank page to starting badly

If you take a single idea from all this, let it be this: you no longer have to fight with the blank page. I used to freeze trying to formulate the perfect request on the first try, and that was exactly the mistake. The way forward is the opposite.

Start badly. Write a clumsy, incomplete request, however it comes out, and let the AI itself correct it with the meta-prompting trick, or apply a template to it, or add an example. The dialogue you learned to sustain in the previous stretch is what turns a bad start into a good result: you don't get it right in one shot, you fine-tune it between the two of you. Asking well stops being an exam and becomes a conversation that starts weak and improves as it goes.

With these shortcuts you already get decent requests out without thinking about theory. What comes from here is learning to shape with more precision how it answers you —the tone it takes on, the way it presents it to you—, so that the answer not only gets the substance right but arrives exactly as you need it.

Definitions

- Prompt: the request you write to the AI. A good prompt doesn't depend on talent, but on a handful of tricks that get copied and reused. - Meta-prompting: asking the AI itself to improve your prompt before answering and to tell you what information it's missing. It takes advantage of the fact that the model knows its workings better than you. - Prompt template: a fixed structure with blanks —role, task, audience, tone, format— that you fill in each case, so you don't have to invent the request from scratch. - Giving an example: instead of describing in words what you want, showing the AI one or more ready-made samples so it catches the pattern and repeats it. - Step-by-step answer: asking it to reason out or work the task in stages before giving the conclusion. It works well when the task is complex and you're using a fast model; reasoning models already do it on their own.

Further reading

- Cursos Femxa, Cómo escribir buenos prompts para ChatGPT: consejos y técnicas — a plain-language collection of tricks, with a role/task/context/format/tone template and the technique of giving examples. https://www.cursosfemxa.es/blog/prompts-chatgpt - Outbounders, Cómo mejorar prompts con IA: método práctico de meta-prompting — explains in Spanish the trick of using the AI itself to rewrite and fine-tune your requests. https://outbounders.es/como-mejorar-prompts-con-ia - IEBS, ¿Cómo escribir mejores prompts en ChatGPT? — practical advice in Spanish: give context, adopt a role, ask step by step and use examples. https://www.iebschool.com/hub/como-escribir-mejores-prompts-en-chatgpt-tecnologia/