AI prompts – what a prompt is and how to build a good one

Definition

A prompt is the instruction you give an AI model – the text that steers how ChatGPT, Claude or Gemini responds. The clearer the role, context, task and format in an AI prompt, the better the answer.

What is a prompt?

A prompt is the text you enter into an AI model to get it to do something: answer a question, write a draft, summarise a document or analyse a table. As an everyday English word, "prompt" means both "immediate" and "a cue to act", and anyone who has worked in a terminal will recognise the command-line prompt – the character waiting for the next instruction. Today, most people mean the AI sense.

Technically, the prompt is the whole instruction sent to an LLM, the model behind tools such as ChatGPT, Claude and Gemini. The model knows nothing about what you want beyond what is in the prompt. That is why two people using the same tool can get completely different results: the difference is rarely the model – it is the instruction.

AI prompts also differ from ordinary Google searches. A search is often just a few words ("web agency stockholm"), whereas a prompt to generative AI is longer and more descriptive – people write whole sentences, explain their situation and ask follow-up questions. That changes how content has to be written to stay visible, which we return to below.

How is a good AI prompt built?

A good prompt has five parts: role, context, task, format and examples. They do not all need to be there every time, but the more that are missing, the more the model has to guess. In practice, the three big providers' own guides – Anthropic, OpenAI and Google – use the same breakdown, even if their wording differs slightly.

• Role – who the model should be ("you are an experienced B2B copywriter").
• Context – background the model cannot know: company, audience, channel, tone.
• Task – exactly what should be done, preferably one thing at a time.
• Format – length, structure, language, heading levels, bullet list or running text.
• Examples – one or two samples of what a good result looks like.

Here is a concrete example. A weak prompt: "Write a LinkedIn post about our new service." The model does not know which service, who will read it, how long the post can be or how you usually write. The result is generic and has to be rewritten.

A better prompt: "You are a marketer at a Swedish accounting firm serving owner-managed companies with 5–50 employees. Write a LinkedIn post of at most 120 words about our new bookkeeping service with a fixed monthly price. The tone should be factual and warm, no exclamation marks, no hashtags. End with a question to the reader. Here is an earlier post that worked well: [paste in]." Same tool, but now role, context, task, format and example are all in place.

Examples often make the biggest difference. Including samples is known as few-shot; describing the task alone is zero-shot. Zero-shot is enough for simple tasks; few-shot works better as soon as style, structure or judgement is hard to put into words.

What is prompt engineering?

Prompt engineering is the systematic process of designing, testing and refining prompts until they deliver reliable results – every time, not just once. It differs from "writing a prompt" roughly as writing an email differs from building a template that a hundred colleagues will use.

A well-known technique is chain-of-thought: asking the model to reason step by step before it answers. The technique was described by Wei et al. (2022) and clearly improved results on arithmetic tasks. Later research adds a caveat: a 2024 review shows that the gain is greatest for mathematics and symbolic reasoning, not for ordinary writing tasks. So prompt engineering is not magic – it comes down to clarity, structure and testing.

A common question is why the same prompt gives different answers in ChatGPT, Claude and Gemini. The models are trained on different data, have different built-in instructions and interpret formatting requests differently. A prompt that is clear on all five parts works best everywhere, but expect to fine-tune it for each tool. The vaguer the prompt, the greater the risk of AI hallucination – the model filling the gaps with made-up facts that sound plausible.

What is the difference between a system prompt and a user prompt?

The system prompt sets the framework for the whole session; the user prompt is the task in each individual turn. The system prompt holds the role, rules, tone and anything that should apply whatever is asked: "Always answer in English", "never refer to competitors", "answer 'I don't know' if the information is missing". The user prompt is then the concrete part: "Summarise this meeting", "Make this paragraph shorter".

If you are just chatting, the difference hardly matters – most tools have a hidden system prompt you never see. It starts to matter when a company builds its own solutions. An AI agent that answers customer questions or sorts incoming email is steered almost entirely by the system prompt; that is where the company's rules, knowledge and limits are set out. Write a good system prompt once and you avoid repeating the same instructions in every user prompt.