How to write one prompt that gets a usable answer on the first try

We ran the pair on 2026-08-12: same model, same machine, same hour, same deliverable. One prompt written the way most people type it. One written the way this article recommends.

The thin version was seven words. “Write the hero copy for my services page.” It came back with 143 words and not one line of copy in it. Three clarifying questions, and a reasonable request to know whose services page this even was.

The specified version was 115 words. It came back with three complete options, first try, and every constraint held.

Worth being honest about one detail, because it’s the interesting part. The thin prompt asked questions rather than inventing an answer, and it did that because it was running in a tool with real context about the machine it was on. Type those same seven words into a plain chat window and you usually get the opposite failure: confident copy for a business that doesn’t exist, in a voice that isn’t yours, which you then have to argue it out of.

Both outcomes cost you the same thing. A round trip you didn’t need to spend.

This is the whole system. Six slots. Fill them and you can stop thinking about prompting as a craft.

Filling a slot badly is the same as leaving it empty, and the reader slot is where you can see that most clearly. “Small business owners” and “a shop owner embarrassed by their site who was just quoted $15k and four months” are both answers to the same question. Only one of them changes a single word of what comes back.

The last two slots are the ones almost nobody writes, and they’re the ones that separate a draft you can send from a draft you have to fix.

Because a gap in a brief is a gap in the output, and the model fills gaps rather than leaving holes. It isn’t lying to you. It’s completing a pattern, and every marketing page it has ever seen has a testimonial on it.

Three things we’ve actually had to catch on our own work, all of them from prompts that were otherwise pretty good:

• A testimonial from a client who never said it. The brief listed which people were approved to quote. It didn’t say “invent nothing else,” so a plausible quote from a plausible person appeared, and it read beautifully, which is exactly what made it dangerous.
• Material that was supposed to stay private. A brief reserved part of the methodology for the sales conversation. That reservation lived in the brief and never made it into the prompt, so the whole thing got published on the page.
• A number baked into the copy. A section headline said “five services.” Adding a sixth meant rewriting the copy and the code, because nobody had said “this list will grow, don’t state the count.”

Every one of those is a one-line fix in the prompt. Write down what may not be invented, write down what’s out of scope, and write down what has to stay flexible. It feels like over-explaining right up until the moment it saves you from shipping a quote nobody said.

The rule that catches almost all of it: tell it to mark what it doesn’t know instead of filling it in. “Where you don’t have a fact, write [NEEDS FACT] rather than guessing.” Then the gaps arrive labeled instead of disguised.

This is the exact prompt from the test above, unedited. It took about four minutes to write.

Write hero copy for the services page of a two-person web studio.

READER: a small-business owner who already has a website they are
embarrassed by, and who has been quoted $15k by an agency and told it
will take four months.

JOB OF THE PAGE: get them to book a free first hour.

VOICE: warm and direct, short declaratives, no hype words, no em-dashes.

CONSTRAINTS:
- Do not invent statistics, client names, years in business, or awards.
- Do not state a number of services; the list will grow.
- Headline under 10 words. Subhead under 30 words. One button label.

OUTPUT: exactly three options, each as headline / subhead / button, and
nothing else.

Notice what isn’t in there. No “you are a world-class copywriter.” No “please” and no “thanks so much.” No three example headlines to imitate. Nothing about the industry’s evolving landscape. Just the six answers, in plain language, in whatever order they occurred to me.

The labels in caps are for you, not for the model. They make it obvious at a glance which slot you left empty, which is the entire reason the format is worth keeping.

The pattern underneath all five: information about you and your reader is worth a lot, and information about the model is worth almost nothing.

Rewrite the prompt. Don’t negotiate with the answer.

The instinct is to reply “make it shorter,” then “warmer,” then “actually go back to the second one.” That path works eventually, and it’s the expensive way there. Every follow-up drags the whole conversation along with it, and you end up steering a draft built on a brief that was wrong from the first line.

So read what came back and figure out which of the six slots it exposes. Generic voice means the reader slot was thin. Wrong length means the form slot was missing. Invented facts mean the facts slot was empty. A wrong answer is diagnostic, and it’s pointing at the line you didn’t write.

Then start clean with the fixed prompt. Fresh conversation, one message. This is the same economics that makes the difference on a build: think it through where thinking is cheap, hand over one finished spec, and get out.

And once a prompt shape works twice, stop retyping it. Save it. When you’ve used the same one three times it has earned a permanent home where your assistant can load it without being asked.

This only works when you know what you want. If you’re still figuring out what the page should even say, you can’t fill the six slots, and pretending you can just produces a confidently specified version of the wrong thing. Have the messy conversation first. Write the prompt once the conversation has produced answers.

It also can’t help with anything that depends on facts the model doesn’t have. Telling it not to invent your revenue number doesn’t give it your revenue number, it just gets you a blank where the number goes. Which is the correct outcome, and still means you have to go get the number.

Longer is not better past the point where every slot is filled. A prompt with real constraints and one clear job beats a two-page brief with contradictions buried in it, and the second one takes far more of your time to write.

And the numbers here come from one pair of runs on 2026-08-12, not a study. The direction has held across every project we’ve run this way, but treat the specific figures as an illustration rather than a benchmark, and go run the pair yourself on a task you actually care about.