What Is a Prompt? Prompts and System Prompts

A prompt is the instruction you give an AI such as ChatGPT or Claude: your question or task, plus the material and the requirements for what the result should look like. This article is about prompts in AI, not the command prompt in a computer terminal, which is also called a prompt. Besides your own prompt there is the system prompt: instructions set before the conversation that apply to every message, for example role, tone and limits.

I write my prompts with a simple four-part structure that I show further down. For recurring tasks I don't type them anew every time: my AI employees keep their standing instructions in files, each in a folder of its own. This article covers the basics: what a prompt is, which kinds there are and what makes a good prompt. All the basics about AI are collected on my AI Basics page.

What a prompt is

The English word prompt means a request or a cue. With an AI, it is everything you give it as input. That includes:

• the task, for example "Draft a reply to this customer",
• the material, such as the customer's email, a price list or a file,
• the requirements: length, tone, format and what the AI should not do.

The AI does not read your prompt on its own. It also sees everything already in the conversation and the fixed instructions in the background. Together, that is called context. How much of it fits in at once is explained in my article on the context window.

Anthropic, the company behind Claude, offers a good picture for this: think of the AI as a brilliant but new employee who does not know your norms and workflows. The more precisely you explain what you want, the better the result.

Prompt, system prompt, project instructions: the terms

A few notes on this:

• The provider writes the chat app's system prompt. Anthropic discloses that the Claude app in the browser and on mobile passes a system prompt at the start of every conversation, for example with the current date. You do not see it in the chat.
• Your own standing instructions go into the settings or into a project. In Claude, according to Anthropic, the instructions in the settings apply to all of your conversations. Project instructions apply only within that project. See the Anthropic projects guide for current availability. There you can set, for example, that Claude answers more formally or from the perspective of a specific role or industry. Put simply, these instructions take on the role of a system prompt for you. How I set up projects is in How to Use Claude Projects Right.
• CLAUDE.md is a text file that the coding tool Claude Code, according to Anthropic, loads into context at the start of every session. So it is a standing prompt for a folder. How it works together with an AGENTS.md file, its counterpart for other coding tools, is in CLAUDE.md vs AGENTS.md: What Survives.

How a good prompt is built

To get started, I use a structure of four building blocks myself:

1. Role and task in one sentence.
2. Context, in the narrower sense here: the material and what the AI needs to know about it.
3. Format: length, structure, tone.
4. Limits: what the AI should explicitly not do.

An office example, made up for illustration. A weak prompt reads: "Write an email about the delay." The AI does not know who is writing, to whom, how late it will be or what you want to offer. It fills the gaps itself, and that is exactly where errors creep in.

A better prompt reads: "Help me explain a delay to a customer. The customer ordered a shelf, the delivery date was November 3, the new date is November 10 because a part is missing. Write in a friendly, concise way, five sentences at most, with an apology and the new date. Do not offer a discount and do not invent any other reasons."

Two tips from Anthropic's guide make the difference even bigger:

• Say why. When you explain a requirement, the AI can apply it better. Anthropic's example: instead of "Never use ellipses", better "Your response will be read aloud by a text-to-speech engine, so never use ellipses since the engine will not know how to pronounce them".
• Show examples. According to Anthropic, examples are one of the most reliable ways to steer format, tone and structure. The recommendation is three to five, as close to your real case as possible.

And a simple test, also from Anthropic: show your prompt to a colleague who does not know the task. If they would be confused, the AI will be too.

When you write a system prompt

In the normal chat app you do not write a real system prompt. There you use the instructions in the settings or projects. You only write a system prompt when you build an assistant yourself, for example through a provider's interface or with a tool for your own AI assistants. Anthropic puts it this way: even a single sentence about the role in the system prompt changes behaviour and tone for your use case.

An example of a short system prompt, made up for illustration: "You answer questions from customers of a carpentry workshop about delivery times and caring for wooden furniture. Answer in a friendly way in two to three sentences. About prices you say: the team will get back to you personally."

What to watch out for

• No customer data in the prompt without thinking. Names, addresses and contract details only belong in it if you know where they are processed. How I anonymise client data beforehand is in the section on it in Claude Prompts: Templates.
• A good prompt does not replace checking. Even with the best request, the AI can make something up. Check figures, dates and promises yourself. How to spot that is in AI Hallucinations: Why AI Makes Things Up.
• Don't retype prompts for recurring tasks. Save good prompts as templates or store them as project instructions. Ready-made briefs to copy are in Prompt Templates: 12 Copy-and-Paste Briefs for Small Businesses.
• You can speak long prompts. If typing is a chore, dictate your prompt. How that works is in Dictating Prompts to Claude and ChatGPT.

Frequently asked questions

What is a prompt, explained simply?

A prompt is your input to an AI: the question or task, the material that goes with it and the requirements for the result. The more precise it is, the less the AI has to guess.

What is a system prompt?

A system prompt is a set of instructions fixed before the conversation that applies to every message, such as role, tone and limits. In chat apps the provider sets it. If you build an assistant yourself, you write it yourself.

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

The user prompt is your single message in the chat. The system prompt comes before it and applies to the whole conversation. Put simply: the system prompt describes the job, the user prompt the specific task.

How do I write a good prompt?

With four building blocks: role and task, context, format and limits. Explain your requirements and, where you can, give an example of the result you want. Then check the answer before you use it.

How long should a prompt be?

As long as the task needs. A short question needs one sentence, a reply to a customer more like five. What matters more than length is whether someone without background knowledge would understand what you want.

Where to go from here

Take the last prompt you were not happy with and rewrite it using the four building blocks. Then compare both results. If you want to give the AI a lot of your own documents, read the article on RAG systems first. For Claude, I have collected my tried-and-tested tips in Claude Prompts: Templates. You will find every technical term from this article in my AI glossary. How others build their prompts and AI employees is what I show in my community AI Employee Framework.