Every summer Italy gets a degree in something. The national team, the spread, the virus. This year it’s AI’s turn.
For those who have already tackled the subject seriously — and if you’ve read this far, the answer is yes — beach debates are a minor form of torture: the right words used at random, the tone of someone who’s understood everything, confidence inversely proportional to knowledge.
So I decided to turn it into a glossary, in the style of “AI without noise” — the newsletter: one term a week to vaccinate ourselves against shallow-end thinking. One term at a time, properly taken apart.
TODAY: prompt engineering.
Let’s start from the first thing you do with AI: you talk to it. The message is called a “prompt.” Writing it well even has a pompous name: “prompt engineering.”
And here’s the interesting part. “Engineering” evokes secret formulas, a thousand magic prompts, spell words: “act as an expert,” “think step by step,” “I’ll tip you $200.” But the trick isn’t there. There are no spells. The engineering isn’t in the right formulas — it’s in actually saying what you want.
Which is the uncomfortable part: we almost never know precisely. If you’re talking to a colleague, you don’t need to — she fills in the gaps herself, with context and common sense. If you talk to AI, it fills in the gaps too — but randomly. Prompt engineering explains how to fill in the gaps before AI fills them in randomly.
It doesn’t mean writing an essay every time. It means choosing the two or three details that shift the result and letting the rest go. Those who skip this step don’t get poor results because of AI: they get their own vagueness back as a reply — and from there to disappointment is a short step.
One piece of good news, for once against the hype: tricks matter less and less. Models improve, they ask for clarifications themselves. Templates go stale in three months. Clear thinking doesn’t.
I wrote a book about it. And if it takes a whole book to explain how to say what you want, the problem isn’t in the AI…
The perfect prompt doesn’t exist. What exists is knowing what you want — and saying it.
What happens when you give AI not just that message, but also tools? Agents.