AI Prompts: The Method to Write Prompts That Get Useful Answers

Writing an AI prompt is a skill close to writing a good brief for a colleague. Model providers publish their own advice, such as the OpenAI prompt engineering guide, and the principles below are shared by most of them. Our glossary entry on prompt engineering covers the vocabulary.

What is an AI prompt?

A prompt is everything the model reads before it answers: your question, the instructions, any pasted document and the earlier messages of the conversation. The model has no other information about your goal, your audience or your standards. A vague prompt gets an average answer, because the model fills the gaps with the most likely generic choices.

What are the parts of a good prompt?

1. Role

Tell the AI who it should act as: "You are a B2B copywriter who writes for finance directors." The role sets vocabulary, tone and level of expertise.

2. Context

Explain the situation: the company, the product, the audience, the goal of the text, what has already been done. Context is the part most people skip and the one that changes results most.

3. Task

State one clear task with an action verb: write, summarize, compare, rewrite, list. Split large jobs into several prompts.

4. Constraints

Length, tone, words to avoid, facts to include, reading level, language. Constraints turn a generic answer into one that fits your brand.

5. Format

Say how the answer should look: a table with given columns, a numbered list, three versions, a JSON object, headings and short paragraphs.

6. Examples

Show one or two examples of what you expect. This technique is called few-shot prompting, after the research paper Language Models are Few-Shot Learners, which showed that large models can pick up a task from a handful of examples placed in the prompt. See our entry on few-shot learning.

Which prompting techniques work best?

• Ask for steps: on reasoning tasks, asking the model to work through the problem step by step improves accuracy. The idea was formalized in research on chain-of-thought prompting.
• Give reference material: paste the source text, data or style guide and ask the model to use only that. It reduces invented facts.
• Ask for a self-review: "Check your answer against the constraints and correct anything that does not fit."
• Ask questions first: "Before writing, ask me the questions you need answered." Useful when you are not sure what context to give.
• Iterate: treat the first answer as a draft and refine with precise feedback ("shorter", "more concrete", "remove the jargon").
• Separate instructions from content: put pasted text between clear markers so the model does not confuse it with your instructions.

Bad prompt vs good prompt: marketing examples

What mistakes should you avoid?

• Too little context: the model guesses your audience and goal.
• Several tasks at once: quality drops on each one.
• Contradictory instructions: "short but detailed", "formal but casual".
• Asking for facts without sources: the model may produce plausible but false figures. Provide the data or verify everything.
• Accepting the first answer: two or three rounds of feedback usually make the difference.
• Pasting confidential data into tools whose data policy you have not checked.

For ready-made SEO and GEO prompts built on this method, see our library of SEO and GEO prompts. To refine prompts over time, our entry on prompt optimization explains how to test versions against each other.

What about the prompts your customers type?

Prompts also matter from the other side. When a buyer asks ChatGPT "which accounting software for a small agency?" or asks Perplexity "best CRM for real estate", the answer names a few brands. Those prompts work like search queries: they are longer, more conversational and more specific, and they increasingly decide who gets considered.

Following a list of these prompts and checking which brands appear is called prompt monitoring. It tells you on which questions you are recommended, on which you are absent and which sources the AI cite instead. That data then guides your content: pages that answer those questions clearly, with sourced facts, are the ones assistants cite.

A good AI prompt gives the model what a good colleague would need: a role, the context, one task, clear constraints, a format and an example; and the prompts your customers type deserve the same attention, because they decide whether the AI recommend you. To track those customer prompts in ChatGPT, Claude and Gemini, Sorank follows 25 to 150 prompts depending on the plan, shows who is cited in your place and publishes optimized articles on autopilot, with a 3-day free trial.