Prompt

A prompt is the text you type into an AI program like ChatGPT. It’s the instruction or question with which you set the program in motion. It can be a single word or an entire page. Images, tables, or program code can also be part of a prompt. Anything the system receives as input belongs to it. The program then computes and delivers a response.

Why this matters

With ordinary software, there are buttons and menus. With AI programs, the prompt is the user interface. Anyone who doesn’t formulate it well gets poor results – even though the model could do more.

The difference is often significant. The question “Explain inflation to me” delivers a generic text. “Explain inflation in five sentences for a 10th-grade class test” delivers something usable. Same model, different input, clearly better result.

That’s why an entire craft has emerged around prompts. It’s called prompt engineering: the systematic trying-out and refining of inputs. Companies keep collections of tested prompts, similar to templates for letters.

How it works

Language models are trained to continue text. They calculate which word fits best next. The prompt is the beginning that the model continues writing. So there’s no database in which an answer is looked up.

From this follows a practical rule: context helps. If the prompt states who will read the answer, how long it should be, and in what tone, the continuation becomes more fitting. Examples within the prompt are especially powerful. If you show the model two finished examples, it imitates their structure.

A prompt does, however, have an upper limit. The model can only take a limited amount of text into account at once – this working memory is called the context window. With very long inputs, the beginning or middle can slip out of view. This is also relevant in chat: there, the entire conversation history so far is sent along with every new message. So the prompt grows with every round.

There’s also a part that users don’t see: the system prompt. In it, the provider defines in advance how the program should behave, for example staying polite and not giving instructions for weapons. This hidden instruction comes before anything you enter yourself.

Where you encounter the term

Wherever AI tools are used. With image generators, the prompt describes the desired subject along with style and perspective. There are websites where users exchange successful image prompts like recipes.

In job postings, prompt competence is now appearing as a requirement, for instance in marketing or software development. Some companies train entire departments in it.

In the news, you often read the term in connection with security. In a prompt injection, someone hides an instruction in a document or on a website. If an AI assistant reads this text, it takes the hidden instruction to be a genuine command. This can be used to bypass security rules or extract data. This is currently considered one of the most difficult unsolved problems with AI products.