Prompt Structure 101: What Actually Makes a Prompt Work

Ask for "a good email" and you get something forgettable. Ask for "a 100-word follow-up to a client who missed a call, professional but warm, ending with a specific date to reschedule" and the same model suddenly delivers. Nothing about the model changed between those two requests - what changed is structure, and structure takes about five minutes to learn.

There are four ingredients that separate a prompt that lands from one that flops: context (who the output is for, what's already going on, any background the model genuinely needs), task (spelled out directly rather than buried in a pile of setup), format (the actual shape of the output - how long, how structured, what tone), and constraints (what to leave out, what has to be there). Weak prompts are almost never missing all four - usually it's format and constraints that get skipped, and that's precisely why results come back generic or off on length and tone even when the substance is basically right.

Breaking the Four Parts Down Individually

It's easier to run through these as four separate checks before you hit send than to try cramming everything into one run-on instruction. Leaving one out won't stop the model from answering - you'll still get something back - but you'll likely end up rewriting the output, which kind of defeats the purpose of asking at all.

• Context: Who's this for, and what background does the model actually need that isn't obvious from the request itself - a client relationship, a product category, an earlier conversation.
• Task: The concrete action, phrased as a verb plus an object - 'summarize,' 'rewrite,' 'draft a reply to' - rather than left for the model to infer from surrounding background.
• Format: Length, structure, tone - word count, bullets versus prose, formal versus casual - spelled out the way you'd brief a new hire, not left for the model to guess at.
• Constraints: What has to stay out (jargon, exclamation points, a certain claim) and what has to be in there (a date, a name, a call to action).

The Same Ask, Rebuilt

Look closely and none of the stronger versions inject extra opinion or creative flair into the ask. What they add is clarity about the shape of the output and what to steer clear of. That's really the whole game - the model already knows how to structure a product description; it just doesn't know your word count, your audience, or your house voice unless you spell it out.

The Single Trick That Beats Everything Else Here

Spelling out a tone in words - "professional but friendly" - is a far blunter tool than just showing the model a sample to follow: "match this style: [paste something short]." If you have anything close to what you're after lying around, an old email, a sentence that clicked with someone, dropping it in does more work than another paragraph of description ever could.

The reason it works: tone descriptors are fuzzy in a way examples never are. 'Professional but friendly' can mean radically different things to different readers, but pasting in one actual sentence kills the ambiguity outright - now the model is matching a pattern rather than interpreting an adjective.

When the Prompt Comes Back Wrong

A missed first draft isn't a failure - it's data. Pin down exactly what's off (too long, wrong tone, a detail that got dropped) and turn that into an explicit constraint for the next attempt. Two or three rounds of this kind of surgical correction generally beats trying to write the one perfect prompt on the first try.

• Diagnose the miss, not just the result: Pin down what specifically went wrong - length, tone, a missing fact - instead of just noting that it felt off.
• Fix one thing at a time: Fold the specific correction into your next prompt rather than rewriting the whole request from zero.
• Keep what worked: If a phrase or structure landed well, carry it into the next round instead of starting from scratch.

Run This Checklist Before You Send Anything

A checklist catches most prompt problems faster than instinct does, at least until the four-part habit becomes automatic. Before firing off anything longer than a one-liner, run it past these four questions.

• Does the model actually know who this is for, or are you assuming it can guess?
• Is the task stated plainly, or is it buried somewhere inside a paragraph of context?
• Have you specified a length, structure, or tone - or is 'good' left completely undefined?
• Is there anything the output has to avoid or include that you haven't actually said?

None of this calls for special jargon or a course in prompt engineering. It's the same discipline you'd use briefing a freelancer or a new hire - be clear about who it's for, what you're after, what it should look like, and what to steer away from - just aimed at a model instead of a person.