How to write a clear AI prompt (with a reusable template)

Write clearer AI prompts with a five-part structure, a worked launch-email example, and a reusable template you can adapt to your own work.

A useful AI prompt tells the model what to do, gives it the information needed to do it, and describes what a good answer should look like. You rarely need a long introduction. You need a clear brief.

This guide gives you a practical structure, a worked example, and a way to improve a prompt without rewriting it from scratch.

Start with a task someone could actually follow

“Help with my launch” leaves almost every decision open. Does help mean planning the launch, editing an announcement, or writing an email? Who will read it? What facts are available?

Start with a verb and a deliverable:

Turn these launch notes into a short customer announcement email.

That sentence establishes the job. The rest of the prompt can explain how to do it.

Both Anthropic’s prompting guide and Google’s prompt-design guidance emphasize clear instructions, relevant context and specifying the response you want. The five-part structure below is a practical editorial framework for putting those ideas to work; it is not a formula that guarantees a correct answer.

Use five parts: task, context, input, constraints and output

These parts answer five different questions. Keeping them separate makes it easier to spot what is missing.

Task: What should the model produce or change?
Context: Who is this for, and what situation matters?
Input: Which notes, text or facts should it use?
Constraints: What boundaries must the answer respect?
Output: How should the result be presented?

You do not need five headings for every small request. A quick rewrite may need only two sentences. The structure becomes more useful when a task has several moving parts.

Here is a reusable starting point:

Task
Create {{deliverable}}.

Context
The audience is {{audience}}. The purpose is {{purpose}}.

Input
Use the material below:
{{source_material}}

Constraints
Use only the supplied facts. Flag missing information.
Keep the tone {{tone}} and the length under {{word_limit}} words.

Output
Return {{output_format}}.

The braces mark details to fill in before using the prompt. They are placeholders, not instructions that every AI tool will automatically resolve.

A worked example: a launch announcement

Consider this fictional brief: a small software company has added PDF exports to its reporting tool. The update is available to existing paid customers on 12 November. There is no discount and no pricing change.

A vague request might be:

Write an exciting email about our new reporting feature.

A more complete brief is:

Write a customer announcement email about PDF exports.

Audience: existing paid customers who create weekly reports.
Purpose: explain the update and invite them to try it.

Facts:
- PDF exports become available on 12 November.
- Customers can export a report from its Share menu.
- There is no pricing change.

Keep the email under 140 words. Use a friendly, direct tone.
Do not invent performance claims, discounts or customer quotes.

Return one subject line, one preview line and the email body.
End with one call to action: open a report and try Export PDF.

The second version supplies the decisions that the first version left open. You should still check the result. An instruction to use supplied facts is a useful boundary, not a guarantee that the model will follow it perfectly.

Show an example when style is hard to describe

Words such as “professional” and “friendly” mean different things to different readers. If the tone keeps missing the mark, add a short example of the style you want.

For instance: “Use wording like ‘You can now export a report as a PDF,’ rather than ‘We are thrilled to unveil a revolutionary capability.’” Keep the example relevant to the task, and make clear whether it is a style reference or a fact to include.

Avoid pasting confidential customer emails merely to demonstrate tone. A short fictional example is often enough.

Improve one thing at a time

Run the prompt with a representative input, then compare the result with the brief. Check the facts, audience, requested sections and length.

When something fails, change the instruction closest to that failure. If the answer contains unsupported claims, tighten the source boundary. If it rambles, specify a shorter structure. If the tone is wrong, add a concise style example.

Changing one part at a time makes it easier to understand what helped. Try the revised prompt with a second input before deciding it is reusable.

Save the version that earned its place

A good prompt library should contain prompts you have checked, not every first attempt. Give the working version a task-based name such as “Launch email — customer update,” add a short note about the input it expects, and save it for the next launch.