Imagine handing a task to a brilliant colleague who just started at your company. They don’t know your business, they’ve never seen your clients, they have no idea how you usually write an email. Nobody told them any of this, and yet you say “fix up this text” and expect them to guess the rest. An AI assistant is in exactly this position, every time you open a new conversation.

Wish versus instruction

“Fix up this text” is a wish: it says what you want in general, but not how to tell whether the job is done. “Make this text suitable for a client who doesn’t know the project yet, half a page maximum, formal but not cold” is an instruction. The difference isn’t the length of the sentence, it’s that the second version says how to judge the result. Who is reading it, how long it should be, what tone it needs: these are three things your new colleague can’t guess, and neither can an assistant.

It’s worth noticing what happens when you leave these things out. The assistant doesn’t stop and ask, unless you explicitly invite it to. It produces something anyway, picking an imagined reader, a length, a tone on its own. Sometimes it guesses right, often it doesn’t, and when it doesn’t the tool looks like it failed while really the instruction was missing.

The three things that are almost always missing

Most disappointing requests are missing one of these three things, and none of them requires technical jargon to explain:

• Who the result is for. A summary for you, who already knows the project, is different from a summary for your manager, who is seeing it for the first time.
• What form it should take. A bullet list, a running paragraph, an email with a greeting and a sign-off, a table with specific columns: four different results even starting from the same material.
• What isn’t acceptable. One concrete counterexample is worth ten lines of abstract explanation. Saying “it shouldn’t sound like an excuse” is vague; saying “avoid phrases like ‘unfortunately it wasn’t possible’” is an instruction you can actually follow.

A real example

You need to put together the agenda for tomorrow’s team meeting, starting from scattered notes taken across two separate calls. The version that comes naturally is “make me an agenda from these notes.” What comes back is reasonable but generic: points in the order you happened to write them, with no sense of what’s urgent and what can wait.

The second version adds the three missing pieces: “This agenda is for my team, who know the project but weren’t on either call. I want a list of at most six points, ordered by urgency, each with one line of context. I don’t want a recap of everything that was said, I want only the decisions we need to make tomorrow.” The result changes completely: six points, ordered, each framed as the question the team needs to answer, instead of a report of what already happened. Same starting material, a different instruction, an output you can use right away instead of rewriting.

When the mistake is yours and when it isn’t

Giving these three pieces of information doesn’t guarantee a perfect result, but it changes the diagnosis when something goes wrong. If you didn’t say who the text is for, what form it should take, and what doesn’t work, and the result disappoints you, the most likely cause is you: you let the assistant choose things you actually knew and didn’t say. If instead you gave all three clearly and the result is still wrong, at that point the problem is probably a real limit of the tool on that specific task, not something you missed. Telling the two cases apart saves you from two opposite mistakes: blaming the tool too soon, or pushing too long on a task the tool simply can’t do well.

Something to try, right now

Take the last request you made to an AI assistant, the one that left you a bit disappointed. Rewrite it adding the three things: who the result is for, what form it should take, one example of what doesn’t work. Send both versions, the old one and the new one, and compare the two outputs side by side. Most of the time the difference is obvious at a glance.

A prompt to start from

[Describe the task in one sentence].
The result is for [who will use it, what they already know and what they don't].
It should take this form: [list, paragraph, email, table — be specific].
It doesn't work if: [one concrete example of a mistake to avoid].
Starting material: [text, notes, data].

Where to go next

When the result is still wrong after you’ve given these three things clearly, the next chapter in the pro track, From prompt to workflow, shows you what to do when the same kind of request keeps coming back: at that point you’re no longer writing a prompt, you’re designing a process.