The Four-Part Prompt Formula Every Beginner Should Learn

The reason most people think AI is 'useless for real work' isn't that the model is dumb. It's that they're handing it half a question and expecting a full answer.

There's a formula every good prompt follows, four parts. Skip any one of them and you'll get the kind of generic, mealy-mouthed reply that makes you think 'eh, this isn't actually that helpful.' Include all four and you'll get something you can actually use.

Here it is.

The formula: ROLE + CONTEXT + TASK + OUTPUT

Every useful prompt has four components, in this order:

• ROLE, who you want the AI to be
• CONTEXT, the situation, constraints, and background it needs
• TASK, what you actually want it to do
• OUTPUT, what format you want the answer in

Most beginner prompts are 100% task and 0% of the other three. That's why the output is generic, because the request was generic.

Before and after

Here's the same question, asked two ways:

Why each piece matters

Role

Giving the AI a role narrows the vocabulary and frame it pulls from. 'You're a small business attorney' produces different language than 'You're a startup founder.' Both might be technically correct; only one matches your context.

Context

This is where 80% of the magic happens. The AI doesn't know what you know. Every detail you'd assume a human collaborator already had, the audience, the brand voice, the constraints, the history, has to be made explicit. Otherwise the model fills in averages. Averages are bad.

Task

Be specific. 'Write me an email' is bad. 'Draft a 4-5 sentence email apologizing for X, offering Y, and inviting them back' is good. Verbs and quantities make the task crisp.

Output

If you don't specify format, you'll get whatever shape the model defaults to, which is often a wall of paragraphs when you wanted a bullet list, or a JSON object when you wanted a plain-text email. State the format you want and you'll get it.

A copy-paste template

Drop this into Claude, ChatGPT, or Gemini, fill in the brackets, and you're done:

You are a [ROLE: e.g. "marketing director for a small B2B SaaS company"].

Context:
- [Audience or who you're writing for/about]
- [Brand voice or tone, be specific, e.g. "warm, plain-language, never corporate"]
- [Constraints, length, what to avoid, anything that's been tried before]
- [Goal, what does success look like]

Task:
[The specific thing you want, with verbs and quantities. E.g. "Write a 4-bullet list of subject lines for a re-engagement email."]

Output format:
[What shape you want the answer. E.g. "Numbered list, each item under 60 characters, no explanation."]

When AI keeps giving the wrong answer

Even with the formula, you'll sometimes get something off. The fix is rarely a longer prompt, it's a sharper one. Three techniques:

Ask the AI to ask you questions

Add this line to any prompt: 'Before you answer, ask me up to 5 clarifying questions so the answer is actually useful.' Suddenly the AI is interrogating you, and you're surfacing context you didn't realize you needed.

Show, don't tell

If you can't describe the voice or style you want, paste an example. 'Write this in the voice of the example below' followed by an actual block of text works dramatically better than 'write it conversationally'.

Iterate, don't restart

If the first draft is 80% right, say 'Take that draft and make these three changes: [X, Y, Z]' instead of starting over. The model already has context, use it.