How to write a prompt the model gets right first time

The model does not ask. It guesses

You write «write me a summary» and you get a summary. The problem is which one: you did not say how long, or who for, or whether you want prose or bullets, or what it should emphasise. The model asks you none of that. It chooses on its own, and because it chooses on its own, it gets it right about one time in three.

What follows is three messages of corrections: «shorter», «in bullets», «focus on the conclusions». That is where the time you thought you saved goes, and each correction drags the earlier context along, so the model has more and more noise to work from.

The answer is not writing enormous prompts. It is leaving nothing important unsaid. What follows are the eight things most often missing, in the order worth checking them.

1 · The goal: what it must do, not what it is about

This is the most common gap and the hardest to spot in your own writing, because in your head the goal is obvious. Describing a situation is not asking for anything.

«We have high churn in our SaaS and we do not know why» is excellent context and contains no instruction at all. Do you want hypotheses? A plan to investigate? An email for the customers leaving? The model will pick one, and it has no way of knowing which you meant.

The check is mechanical: look for the imperative verb. If your prompt has none — write, summarize, compare, classify, fix — then you have not asked for anything yet.

2 · The output format

What you are going to do with the answer decides the shape it needs, and only you know that. A summary for your own eyes is nothing like one going into an email, or one your code is going to parse.

If the answer goes to a program, this stops being a preference: saying «return only JSON with the keys title and summary, with no text around it» is the difference between working and getting a «Sure! Here you go:» in front of the JSON that breaks your parser.

• ·Prose, list or table.
• ·With or without headings.
• ·Rough length, if it matters.
• ·If a machine will read it: the exact format, and «nothing else».

3 · The constraints: what it must not do

Constraints are the part most often forgotten because they are negative: you think about what you want, not what you do not want. And they are exactly where a model drifts.

«Professional but warm, and it should not read like an advert» is a tone constraint that changes the result completely. «Do not make up data that is not in the text» is another, and it prevents the most expensive problem.

There is a balance here: one constraint too many straitjackets the model, one too few leaves it loose. The practical rule is to state the ones that would make you reject the answer if they were broken.

4 · The context and who it is for

The same request changes entirely depending on the audience. A text about your product for an investor, for a customer and for your engineering team do not even share vocabulary.

And the context you take for granted, the model does not have. If your product has a name that does not appear in the prompt, it will write «your product» or invent one.

5 · The words that can mean two things

«Improve this text» can mean fixing the spelling, rewriting it entirely, or making it shorter. «Analyse this» can mean a summary, a critique or a table of figures. «Make it professional» can mean formal, technical, or simply free of typos.

You do not need to remove all ambiguity from the language: it is enough to pin down the two or three words the result depends on. If reading your prompt suggests two valid and very different answers, that is the word to nail down.

6 · Repeated instructions

Asking twice for the same thing, in different words and in different paragraphs, reinforces nothing. What it does is give two slightly different versions of the same instruction, and the model has to decide which one to follow.

It happens most in prompts that have been edited over time: a clarification gets added at the end without removing the original sentence above. If you say the same thing twice, keep the more precise version and delete the other.

7 · Structure, once the prompt is long

In a three-line prompt structure does not matter. Past a long paragraph it starts to, because the model has to work out what is context, what is instruction and what is the material to work on.

Separating those three — even with a line break and a word in front — avoids the classic failure: the model treating a sentence from the text you pasted as an instruction.

8 · The role, with less enthusiasm than it usually gets

«Act as a marketing expert with twenty years of experience» is the most repeated technique and the one that changes the result least. Current models already answer competently without being asked.

Where it does help is when the role defines the point of view rather than the quality: «review it the way opposing counsel would» changes what the model looks for. «Act as an expert» changes nothing, and takes up room.

The rule that governs the rest: flag what you do not know

While completing a prompt there is a temptation to fill the gaps with something reasonable. It is the most expensive mistake, and it is also the one made by tools that rewrite prompts automatically.

If you never mentioned length and the prompt ends up saying «max 200 words», someone decided for you and you did not notice. The answer will satisfy a requirement you never set, and you will not know why it came out short.

The right move is to leave the gap visible: «Length: [specify]». A bracket is a question still open, and it shows. An invented figure looks like a decision already made.

It is the rule Prampt applies to itself: where the original did not say something, the result carries a bracket. Across the 28-prompt benchmark it is measured against, preservation of the original intent scores 84.6 out of 100 — and that is the number that matters, because it measures how much was NOT invented.

How to tell whether it worked

The test is not that the prompt looks more professional. It is that two different people reading it understand the same request. If your prompt admits two reasonable readings, the model will pick one, and it will not necessarily be yours.

And there is an opposite signal worth watching: if completing it produced requirements you do not remember asking for, they do not belong. A well-written prompt is exactly what you meant, said without gaps — not a more ambitious version of what you meant.