How to write a prompt that gets a usable answer

A prompt that works is not a magic phrase. It is a short brief, and it contains five things: the task, the context, the constraints, the format and the audience. Leave one out and the model fills the gap with an average of everything it has read, which is why "write me an email about the invoice" comes back as polite wallpaper. The skill is not memorizing clever wording. It is writing the brief you would hand a capable freelancer who has never met you, then improving it once you see what comes back.

The five parts, and what each one prevents

You do not need to label these parts, although labels help once a prompt runs past a paragraph. You just need the information to be present. Here is what that looks like in three very different jobs.

Upgrade one: the email you have been putting off

Before: Write an email to a client about a late invoice.

That returns four paragraphs that apologize on your behalf, mention no amount, name no date, and end with "please let us know at your earliest convenience." Nothing in it is wrong. Nothing in it is usable either.

After: You are writing as the owner of a two person design studio. Write an email to Dana at Harbor Foods about invoice 1042 for $3,400, due 12 March 2026 and now 18 days late. We have worked with them for two years and want to keep them. Ask for either payment or a firm date, offer to resend the invoice, and say that work on phase two pauses at 30 days. Under 120 words, warm but direct, and do not apologize for chasing.

The second version produces something you can send after one read. The amount and the invoice number appear, there is exactly one ask, the consequence is stated without threat, and the length forces the model to cut the filler it would otherwise pad with.

Upgrade two: the research summary that says nothing

Before: Summarize the current state of electric vehicle charging.

You get a tour of the topic: encyclopedia tone, no numbers you can act on, and a closing line about how the landscape continues to evolve.

After: I run operations for a delivery company with 40 vans in one city. Compare depot charging with public fast charging for a fleet that size. Cover upfront cost drivers, running cost per mile, overnight versus midday charging time, and the failure modes people underestimate. Give me six bullets of no more than 30 words each, then one paragraph on the single decision that matters most. Where the answer depends on local electricity pricing or grid connection rules, say that instead of guessing.

The audience line ("40 vans in one city") does more work than anything else in that prompt. It rules out advice for private drivers and advice for national fleets in one stroke. The final sentence does not make the model honest, but it does make its guesses visible, which is the part you can then go and check. Anything it states as fact still needs checking against a real source before it reaches a decision.

Upgrade three: the spreadsheet formula

Before: Write a formula to calculate the total.

After: In Google Sheets, column A holds a date, column B holds a category as free text, and column D holds an amount in dollars. I want a formula for cell F2 that totals column D where column B is exactly Travel and column A falls in the quarter beginning on the date in cell F1. Blank amounts count as zero. Give me the formula on one line, then one sentence per argument explaining what it does.

Formula prompts fail for a specific reason: the model cannot see your sheet. Every column letter, every exact category spelling and every edge case you do not mention is a guess it has to make. Naming the tool matters too, because Excel, Google Sheets and Apple Numbers disagree about function names and argument separators.

The second prompt is where the quality comes from

Almost nobody writes the perfect first prompt, and treating the first answer as the finished product is the single most common way people underuse these tools. The first answer is a draft that shows you what your brief actually said.

So read it and name the defect precisely. "Make it better" achieves nothing. "Cut paragraph two, keep the deadline sentence exactly as written, and rewrite the rest for someone skimming on a phone" achieves a great deal. Useful repair lines sound like this: too formal, drop the greeting; you invented a discount, remove it; give me three options instead of one recommendation; that is 240 words, cut it to 120 without losing the deadline.

One caution. If the conversation has gone three or four rounds and the drafts are circling, open a new chat and paste an improved prompt instead. When the job is genuinely large, running it as a few smaller prompts in sequence beats one prompt that tries to do everything at once. A long thread carries every rejected version along with it, and the model keeps drifting back toward them. That failure and eight of its relatives are catalogued in the prompt mistakes that ruin an answer.

A 60 second checklist before you press enter

1. Have you said what you want produced, in one verb? Write, compare, list, rewrite, extract.
2. Have you pasted the facts it cannot know: names, numbers, dates, the source text?
3. Have you set a limit it can count, such as 120 words or six bullets, rather than asking for "brief"?
4. Have you named the format you want back: a table, a list, a paragraph, plain text with no headings?
5. Have you said who reads this and how much they already know?
6. Have you said what to leave out, such as no preamble, no apology, no invented statistics?

Five of those six take one clause each. That is the whole difference between a prompt that saves you 20 minutes and one that costs you 10. A prompt that clears all six is worth saving where you can reuse it.

What to do next

Take the task you gave a chatbot most recently, rewrite it with the five parts, and run both versions in separate chats so the difference is visible side by side. Then pick up the two techniques that extend this structure furthest: showing examples instead of describing what you want, and naming the exact output format so you stop reformatting answers by hand.