How to Write a Prompt That Gets a Useful Answer

A bad answer is usually a symptom of a prompt that left the model to guess. Five specific additions fix most of it, and none of them requires knowing anything about how the model works.

The short answer

A useful prompt supplies five things: who the answer is for and what happens to it next, the actual material rather than a description of it, the shape the answer should take, one example of what good looks like, and an explicit list of what the answer must not do. Most disappointing answers are missing three of the five.

None of this is about magic phrasing. There is no word that unlocks a better model. What these five moves do is remove the guesses: every gap you leave, the model fills with the most statistically ordinary option, which is why unspecified prompts come back reading like a press release nobody asked for.

Move 1: say who it is for and what happens next

“Write a summary of this report” has no destination. “Write a summary for our finance director, who will decide from it whether to approve the spend” has one, and the two summaries are different documents. The audience determines what may be assumed, and the next step determines what has to be in it.

The most useful half of this move is the second half. An answer that will be pasted into an email needs no heading. An answer that becomes a slide needs six words a line. An answer that goes to a lawyer needs the caveats left in. Say which.

Move 2: give it the material, not a description of it

The single largest quality difference in ordinary use is between prompts that contain the source material and prompts that refer to it. A model has no access to your files, your inbox, your company or yesterday’s conversation unless those things are in front of it. If you write “based on our pricing policy”, it will invent a pricing policy, and the invention will be fluent.

Paste the email you are replying to. Paste the paragraph you want rewritten. Paste the four numbers. If the material is long, that is a different technique — see working with your own documents — but the principle is unchanged: it can only work with what is in the window.

Move 3: specify the shape of the answer

Length, format and register, in that order of importance. “Under 150 words” is a real instruction; “concise” is not. “Plain paragraphs, no bullet points” prevents the default bulleted sprawl that makes AI-written documents recognisable at a glance. “The register of an internal note between colleagues who know each other” beats “professional”, which produces the register of a company website.

Word counts are followed approximately rather than exactly. Treat them as a target that moves the answer to roughly the right size, not as a contract.

Move 4: show one example of good

One example is worth several paragraphs of description, because it carries all the things you would not think to specify: sentence length, how formal the greeting is, whether you use the person’s first name, whether you sign off with your job title. This is the cheapest quality improvement available and almost nobody does it.

It does not have to be an example of the same task. An email you wrote last month on a different subject still carries your voice. If the only example you have is of something bad, that works too: show it and say what is wrong with it.

Move 5: name what it must not do

Negative instructions are followed less reliably than positive ones, but they are still worth writing, because the failures they prevent are the ones that make the output unusable. The list is usually short and it is usually the same list every time for a given kind of work.

• No invented facts. “Use only the information I have given you. If something is missing, write [MISSING] rather than filling it in.” The bracketed marker matters: it gives you something to search for.
• No apologising, no throat-clearing. Cuts the opening paragraph that restates the question.
• Named phrases you never want to see. Whatever your own list is. Writing it down once and reusing it is most of what a personal prompt library is.

All five on one request

The task: chase a supplier who has not shipped an order. Here is the version most people write.

Write an email to my supplier about the late delivery.

That produces a competent, generic, slightly grovelling email with a placeholder for the order number and no idea what outcome you want. Here is the same request with the five moves applied.

Write an email from me to a supplier we have used for three years.

CONTEXT (use only this; do not add facts):
- Purchase order 4471, 200 units of the M6 bracket.
- Promised 14 July. Still not shipped on 28 July.
- Second late order this year; the first was in March.
- Our production line stops on 4 August without it.
- I want the stock, not a refund or a credit.

AUDIENCE: Ana, the account manager I deal with directly. First-name
terms. She may forward this to her operations manager, so it has to
read reasonably to somebody who does not know me.

WHAT HAPPENS NEXT: I need a written shipping date by end of Thursday
so I can decide on Friday whether to buy the shortfall elsewhere.

DO:
- Ask for a shipping date in writing, by Thursday.
- State the 4 August production stop as a fact, not as a threat.
- Offer partial shipment of 80 units this week as an acceptable outcome.

DO NOT:
- Threaten to cancel the contract or mention lawyers.
- Apologise for chasing.
- Use the phrase "as per my last email" or "circle back".
- Invent anything not in the context above. Write [MISSING] instead.

SHAPE: under 150 words, plain paragraphs, no bullet points, no subject
line longer than eight words.

EXAMPLE OF MY REGISTER (a different email I sent last month):
"Ana - the March order arrived Tuesday, all counted and fine. One
thing: two of the cartons were open at the corner. Not a problem this
time, but worth a word with whoever packs them. Thanks - Tom"

The second prompt takes ninety seconds to write and returns something you send with one edit. The first returns something you rewrite from scratch, which is why people conclude the technology does not work for them.

The second prompt matters more than the first

When the answer is wrong, the instinct is to rephrase the original request and try again. That throws away the useful information you now have, which is a concrete example of the failure. Say what was wrong with the answer instead.

1. Name the specific defect: “the second paragraph invents a contractual penalty we do not have”, not “make it better”.
2. Say what to keep. Otherwise the repair often destroys the parts that worked. “Keep paragraphs one and three exactly as they are.”
3. If the same defect survives two repairs, it is in the prompt, not the answer. Start a new conversation with the defect written into the DO NOT list. Repairing inside a long thread gets less effective as the thread grows, which is a real and measured effect.

Where this stops working

No prompt makes the model know something it was not given. If you ask about your own company’s policy, a person’s current job, or last week’s news, you will get a confident answer assembled from what was statistically likely rather than from any record. The mechanism behind that is worth ten minutes: why models produce confident false answers.

A false premise in your prompt is accepted rather than challenged. Ask “why did our Q3 revenue fall” when it rose, and you will get four reasons why it fell. Adding “if the premise of my question is wrong, say so first” helps, though it does not fix it entirely.

And contradictory constraints do not produce an error message. If you ask for under 150 words, six specific points and full context for a stranger, something is silently dropped. When an answer keeps ignoring one of your instructions, check whether it can be obeyed at the same time as the others.