How to write a prompt that works

A prompt is a brief for a very fast, very literal contractor. If the brief is vague, the work is vague. Every strong prompt answers five questions, in this order.

1. Who is the model supposed to be?

Open with a role: “You are a financial analyst”, “You are an experienced editor”. A role sets vocabulary, depth, and judgment in one line.

2. What does it need to know?

Context is the part people skip. Audience, goal, source material, what you have already tried. A large language model only knows what you give it plus its training; it cannot see your situation. Paste the relevant facts rather than alluding to them.

3. What exactly is the task?

One task, stated as an instruction: “Produce an outline”, “Compare these two options”, “Extract every action item”. If you have three tasks, write three prompts, or number them explicitly. Burying tasks in a paragraph is the fastest way to lose one.

4. What are the constraints?

Length, tone, format bans, and, crucially, what not to invent: “If a fact is not in the notes, say so instead of guessing.” Constraints are where you prevent the failure modes in our spotting AI mistakes checklist.

5. What shape should the answer take?

A table, a bullet list, JSON, a 100-word summary. If you do not specify the shape, you get the model’s default essay, and you will reformat it by hand.

Worked example

Weak: “write something about our product launch”.

You are a B2B product marketer.
Context: We launch [product] on [date] to [audience]; the one-line pitch is [pitch].
Task: Draft the launch announcement for our blog.
Constraints: 300 words, plain English, no buzzwords, do not invent features or customers.
Output: Title, one-sentence hook, then three short sections.

Related resources

• Prompt templates: this structure, pre-filled for four common jobs.
• Anatomy of a good prompt: the one-screen version of this guide.
• How to fact-check AI answers: for after the model replies.