How to write a prompt for AI: the formula, 10 templates and examples

A prompt is the instruction you give an AI model before it writes anything for you. It describes what needs to happen — not what the model should invent on its own. That distinction looks small, but it decides the outcome: the model doesn’t read between the lines or fill in what you left out.

A short, vague prompt — "write something about coffee" — hands every decision to the model: topic, tone, length, structure. Sometimes it guesses right, often it doesn’t. A prompt broken into parts — who’s reading, what matters, which tone, how long — leaves nothing to guess, and the first draft lands closer to what you actually wanted. The gap between a weak prompt and a strong one isn’t length. It’s how many decisions you made yourself instead of leaving to chance.

How do you write a good prompt without studying prompt engineering? Use a seven-part formula: role → task → context → audience → format → constraints → example. Not every part is needed every time — a quick edit only needs two or three — but the harder the task, the more of these are worth locking down upfront instead of fixing the output afterward. Tone and style aren’t a separate part of the formula; they usually live inside role or constraints, which is exactly where it’s easiest to spell them out in words if a tool doesn’t give you a dedicated setting for them.

• Role. Who the model is writing as — a blog author, a copywriter, an editor. The role sets the vocabulary and the level of confidence the model writes with.
• Task. What exactly needs to happen — write, shorten, rewrite, brainstorm options. One task per prompt, not three at once.
• Context. Where the text will live and what it’s about — the product, the situation, the occasion. Without context, the model invents details instead of using yours.
• Audience. Who’s reading and what they already know. A text for a beginner and one for a specialist need different structures even on the same topic.
• Format. Article, social post, email, a list of headlines — the shape decides the structure of the answer as much as the content does.
• Constraints. Length, what not to mention, which words to avoid. Constraints save you exactly the revision round you’d otherwise spend saying "cut this."
• Example. A short sample of the tone or style you want. Not required, but it settles half the arguments about what "friendly" actually means.

Here’s the formula built up in one pass. Start: "write a blog post about choosing a standing desk" — there’s a task, and the model fills in everything else. Add audience and context: "...for remote workers setting up a home office on a budget." Add role and format: "write as the author of a home-office blog, 600 words." Finally, constraints and an example: "don’t name specific brands, keep the tone friendly and free of jargon — like in how to write an article with AI." Each added part removes one fork in the road the model used to decide on its own.

Ten prompt templates for common writing tasks, below. Each one is a working prompt example with variables in square brackets — swap them for your own details and use it as is. Where the platform has a dedicated tool for the job, there’s a link to it.

Every template above works as is or as a starting point: drop your own variables into the text generator — the brief field takes any of them without reformatting.

The same request — "write about our café’s grand opening" — can be phrased as a prompt at three levels of readiness. The difference isn’t length; it’s how many decisions the author made instead of leaving them to the model.

• Weak. "Write about our café opening." No audience, no tone, no format. The model picks everything itself, landing on a generic text that would fit any café and therefore doesn’t quite fit this one.
• Medium. "Write a social post about our new café on Main Street, for social media, in a friendly tone." Platform, tone and occasion are set, but there are no constraints: it’s unclear whether a call to action is needed, whether to mention opening deals, or how long the post should be.
• Strong. "Write a Telegram post about the café opening on Main Street. Audience: people in the neighborhood. Tone: friendly, no more than one exclamation mark per sentence. Under 500 characters. Mention the hours and one reason to stop by in week one. End with a question to readers." Task, audience, tone, format, constraints and an example of behavior are all covered in a single request.

The first reply is rarely final, and that’s normal — a prompt sets a direction, it doesn’t guarantee a bullseye on the first try. A follow-up message should target one thing at a time — tone, length, or one specific paragraph. The more precisely you point at what to fix, the faster the model fixes it.

A follow-up works better when it names a specific spot: "make the second paragraph shorter" beats "make it better" — the second one is a judgment, not an address, so the edit that follows is a guess too. Can you give the model a sample text to follow? Yes, and it’s often worth it: a short example of the tone you want settles arguments about wording faster than describing it. Keep each follow-up separate — three fixes in one message get applied in whatever order the model picks.

Why doesn’t a model follow the prompt you gave it? Usually the request, not the model, is the problem. Here’s what most often gets in the way of a first-try result.

• Task with no context. "Write about our discount" with no occasion, product or platform — the model invents the context itself, and it rarely matches yours.
• Several tasks in one prompt. "Write an article, then three social posts based on it" — the model’s attention splits across tasks and does a worse job on each.
• No format. Without saying "article," "post" or "list," the answer comes out generic — built for no platform in particular.
• No constraints. Without a length or a list of what to avoid, the model either writes too much or touches something you didn’t want mentioned.
• Contradictions inside the prompt. "Keep it short but cover every detail" forces the model to pick which instruction wins, and its pick won’t always match yours.
• A prompt that runs too long. A paragraph with ten different requirements reads to a model the way it reads to a person — some of it gets lost along the way.
• Politeness instead of specifics. "Could you please kindly write" doesn’t hurt, but it doesn’t help either — the result depends on the facts in the prompt, not how politely they’re phrased.

You don’t need to spell out the whole formula by hand: some of the fields in the iBro text generator already cover part of a prompt for you. Tone and style are dedicated settings, not words you type in — pick an "expert" tone and a "persuasive" style and you get the same result a written-out tone description would.

Length in the form isn’t a word count — it’s "shorter" or "longer": instead of asking for "600 words" inside the brief, set the direction once in the settings. The social format switches the output’s structure for Telegram or Instagram on its own, and formatting — paragraphs, lists, bold emphasis — is a toggle too, not an instruction you write into the text.

The brief field itself (up to 6,000 characters) is where the rest of the formula goes: role, task, context, audience, constraints and, if you need one, a short example — exactly the parts the form has no dedicated control for. If you want emoji in the result, flip that toggle instead of asking for them in the brief.

The formula in one line: role, task, context, audience, format, constraints and an example — not every part is needed every time, but the harder the task, the more of them are worth locking down upfront. The ten templates above drop into place once you swap in your own variables. Once the formula’s assembled, write it in the iBro text generator — tone, style and length are already handled by the form, so you don’t have to spell them out in words.