How to write an AI prompt: the five parts that do the work
Role, context, task, format and check. Add all five and the answer improves at once.
Most people type one line into an AI tool and get one generic answer. The model is rarely the problem. The prompt left out what the model needed. A prompt with five parts fixes that.
Why short prompts give generic answers
A large language model writes the most likely answer to the text you give it. If you write "write a LinkedIn post about my course", the most likely answer is the average LinkedIn post. The model does not know your reader, your offer or your voice, so it guesses.
Each detail you add removes one guess. That is all prompt engineering is. You tell the model what a skilled assistant would ask you before starting the job.
The five parts
- Role. Say who the AI works as, and who the work is for. "You are an editor for a business newsletter. The readers are managers aged 40 and over."
- Context. Give the facts. Your situation, your audience, your source text, your voice. Paste the real document. Do not describe it.
- Task. Ask for one thing, with a clear verb. Rewrite, summarize, compare, list, draft.
- Format. Say what the answer looks like. Length, structure, tone, and what to leave out.
- Check. Tell the AI how to test its own answer before you see it. This one line removes many errors.
The order matters less than the content. Put the role first and the check last. A model follows the start and the end of a prompt most reliably.
A full example
This prompt has all five parts. Copy it, change the details and paste your own draft at the end.
You are an editor for a business newsletter. The readers are managers aged 40 and over who want to start a consulting business. I run a one-person consulting business. Below is my rough draft of this week's issue. My voice is direct and warm, with short sentences. Rewrite the draft so that each paragraph makes one point. Keep my examples and my opening line. Return about 400 words in plain paragraphs, with one subheading per section. Before you answer, confirm that no sentence is longer than 25 words and that you added no new facts. List anything you were not sure about. [paste the draft here]
The part most people skip
The check line is the one that changes results the most. A model can produce a hallucination, which is a confident answer that is false. When you ask it to confirm specific rules and to list what it was not sure about, it catches part of its own errors, and it shows you where to look.
You still read the result. The check line makes that review faster.
Build your next prompt in two minutes
The AI Prompt Builder builds a prompt for you. It now uses a goal-first method for the newest reasoning models, and a separate method for local models. When a prompt works, save it. A set of tested prompts for the tasks you do each week is one of the most useful things a solo business can own.
To go one step further, put the role and the context into a system prompt, so you do not type them again.
A good prompt is the first step. Read the four steps past the chat level of AI to make your prompts run on their own.