Why isn't "just asking" enough?
Because the model does not know what you consider a good answer. You have it in your head — who the letter is for, how long it should be, what it must not say — but the model only sees what you wrote down. Whatever you left out, it guesses.
The Prompt, System Prompt, Temperature article listed four elements; here there are five, because in a good instruction the example and the escape hatch are separate items. You do not need all five every time. But when an answer is wrong, one of them is almost always missing.
1. Task — what it should do
One sentence, one verb. Not a topic but an operation: summarise, list, rewrite, compare, find.
The most common mistake is not a vague verb but two tasks in one sentence. "Analyse the contract and write a client summary from it" is two different jobs with two different right answers. The model attempts both and both come out half-finished. Split them.
2. Material — what it should work from
If you hand over nothing, the model works from memory, and that is where hallucination starts. If you hand over the email, the contract, the spreadsheet, it reads.
Background belongs here too — the things it cannot know by itself: who the client is, what happened so far, what the goal is. One sentence is often enough. "This client has been with us for eighteen months and we are now raising prices" produces a different letter.
3. Format — what it should look like
This is the cheapest fix you can make to a prompt, because it is one sentence and it makes the output immediately usable.
State the form (paragraphs, bullets, table, JSON), the size (how many points, how many sentences, how many words), and if it is a table, the columns. "As a table, columns: item, deadline, owner" — that is all.
If the result is not for you to read but to process further, Structured Output shows how far you can take this. One thing worth knowing: the model does not hold a word count precisely, because it does not know in advance where it will end up. Structure ("five points", "three columns") it handles reliably.
4. Example — show it, don't explain it
If the format or the tone is hard to write down as a rule, give a sample. Two good examples are worth more than ten sentences of explanation, and the providers' own guidance says the same: three to five, different from each other, including one hard case.
It matters that the examples are clearly separated from the instruction, otherwise the model mixes them up with the task. Few-Shot covers this properly.
5. Escape hatch — what to do when it cannot
This is the element people leave out most often, and it is the one that prevents the most damage.
By default the model has no acceptable output for "I don't know", so it guesses instead. Give it one and it will use it:
If the material does not contain the answer, say that it is not in there. Do not invent data.
The same goes for uncertain parts: "where it is ambiguous, flag it, don't decide for me." Rewriting prohibitions is covered in Constraints That Stick.
One task, two versions
Suppose you have to answer an incoming complaint.
Bad:
Write a reply to this complaint email.
This contains exactly one element, the task. The model invents a tone, a length, and — worst of all — it may invent a compensation offer, because that is what tends to appear in complaint replies.
Good:
Draft a reply to the complaint below.
The email: [text of the email] What you need to know: the client has been with us eighteen months, the delay was our fault, and we are crediting the shipping fee. We are not offering compensation.
Format: at most 150 words, three paragraphs — acknowledgement, what we are doing, by when. Tone: direct and factual, apologise once, not three times. If something needed for the reply is not clear from the email, ask me rather than inventing it.
Same model, same email. The second version gives you a usable draft, the first gives you a guess.
Two misconceptions
"The longer the better." No. A long prompt is good because it contains many necessary elements, not because it contains many words. Filler ("this is very important", "be professional") improves nothing — a concrete format request does.
"Politeness helps." "Please" and "thank you" neither improve nor harm the result. Write however is comfortable — just don't mistake it for a prompting technique.
What to take away
If you want ready-made templates to start from, all of 20 Office Prompts follow these five elements. And when an answer is wrong, don't rewrite the whole thing. Go through the five elements and find which one is missing: you didn't state the task precisely, you didn't hand over the material, you didn't ask for a format, you didn't show an example, or you didn't allow it to say no. The fix is almost always one sentence.