Introduction

Most people who struggle with AI prompts share one common problem: they do not know where to start.

They open the tool. They stare at the blank input field. They type something vague, get an output that misses the mark, and assume the problem is the AI.

In most cases, the problem is the starting point - not the tool.

This article walks you through how to write a prompt step by step, starting from a blank page. Each step adds one element to a single working prompt. By the end, you will have a complete, structured prompt and a repeatable process you can apply to any task.

The approach is drawn directly from the seven-element framework introduced in Chapter 4 of Learning Prompt Engineering. No prior experience is required.

Why Building in Steps Works

Writing a prompt all at once tends to produce the same result as writing a paragraph without thinking about its structure first. Something comes out - but it is often unfocused.

Building element by element forces a different habit. You make one decision at a time. You ask whether each element is needed before you add it. By the time the prompt is complete, every part of it has a reason to be there.

This is not a longer process. Once the habit is formed, it takes seconds. What it gives you is a prompt that is easier to test and easier to improve - because you know exactly which part to adjust when the output is not quite right.

The Task You Will Build

This walkthrough uses one realistic task throughout:

Every step below adds one element to this prompt. At the end, you will see the complete assembled prompt with every element labeled.

Step 1 - Task: State What You Want the Model to Do

Start here. Every prompt begins with the task.

The task tells the model what job to do. It is the only element that is always required. Without it, no other element can anchor the output.

Use a specific verb. Vague instructions - help, write something about, do something with - leave too much open. Specific instructions - summarize, list, compare, draft - give the model a defined job.

At this point, the prompt is minimal. It will produce output - but that output may vary widely in length, style, and focus. The remaining steps narrow it.

Step 2 - Role: Frame the Model's Perspective

Ask yourself one question before adding a role: Does a professional identity or frame of reference meaningfully improve this output?

For this task - communicating technical content to a non-technical audience - the answer is yes. A communications coordinator writing for a non-technical team will typically approach vocabulary, register, and detail level differently than an unframed model given the same instruction.

Role is not a default opener for every prompt. Add it when it genuinely shapes how the task should be approached.

Step 3 - Context: Give the Model What It Needs to Know

Context tells the model who the audience is, what the purpose is, and what conditions apply. Without it, the model fills in the blanks on its own - and those assumptions may not match yours.

For this task, the relevant context is: the audience has no engineering background, the update covers three areas (progress, blockers, next steps), and the summary will appear in a team email.

Step 4 - Format: Specify How the Output Should Be Structured

Without a format instruction, the model produces output in a shape it selects. That shape may not match your intended use.

A bulleted list is required here. Three bullets - one per update area. Specifying this directly prevents the model from producing a prose paragraph or a numbered list when a bulleted format is needed.

Step 5 - Tone: Name the Voice or Register

Tone tells the model what voice or register to use. It is distinct from a role. Role defines who is producing the output. Tone defines how it should sound.

The model's default register is often professional and moderately formal, which is adequate for many tasks. But when the task requires something specific, naming it tends to produce more accurate results.

For this task, the tone should be plain and direct. The audience is non-technical. Dense or formal language is not appropriate.

Step 6 - Constraints: Set the Limits

Constraints tell the model what to avoid, what to exclude, and what limits to stay within. Without them, a model asked to summarize a document may produce a longer or more detailed response than the task requires.

This task has two specific constraints: each bullet must be one sentence or fewer, and technical jargon must be avoided. If a technical term must appear, it should be explained briefly in plain language.

Step 7 - Examples: Show the Pattern If Needed

An example shows the model the desired output pattern. It is most useful when the required style is easier to demonstrate than to describe - or when consistent phrasing across multiple outputs matters.

You do not need a long example. A single sentence in the right voice is often enough.

For this task, a short style model helps it match the expected register and sentence length.

The Complete Prompt

Here is the finished prompt with all seven elements labeled:

This is not the only valid version of this prompt. Depending on your task, you may not need all seven elements. But this walkthrough shows what each element contributes - and gives you a complete starting point to adapt.

What to Do If the First Output Is Not Quite Right

A first attempt is not a verdict. It is information.

If the output misses the mark, use the draft - test - refine loop:

1. Identify the gap. Which element is responsible? Output that is too long usually points to a missing or unclear constraint. Output in the wrong style often points to tone or examples. Output that misses the audience usually points to a lack of context or role.
2. Adjust one element. Change the element you identified. Do not rewrite the whole prompt - changing multiple things at once makes it harder to know what produced the improvement.
3. Test again. Compare the new output to the previous version. If it is closer, you have found the adjustment. If not, check the next most likely element.

In many cases, one targeted adjustment - adding a constraint, including a short example, or clarifying the role - is enough to bring the output into range.

Key Takeaways

Related Articles