Prompting

Programs like ChatGPT aren’t controlled via buttons and menus, but through text. You write what you want into an input field and get a response back. This input is called a prompt, the English word for a request or cue. Prompting, then, is the activity of formulating such inputs. That sounds trivial, but it isn’t: the same question can yield useful or useless answers depending on how it’s phrased. That’s why a whole craft has developed around this activity, complete with techniques, rules, and rules of thumb.

Why phrasing changes so much

A language model has no understanding of what you actually want. It only has your text. Everything you don’t write down, it has to guess. If you ask “Write something about nuclear power,” you get some arbitrary text of arbitrary length in an arbitrary tone. If you specify the target audience, length, and purpose, the result becomes noticeably more precise.

This also matters economically. Companies build their products on top of third-party AI models they didn’t train themselves. Their only influence on the model’s behavior often runs through the prompt. Behind many chat assistants, then, lies a long, invisible system prompt. It contains instructions like “Respond politely” or “Don’t give medical diagnoses.”

For a while, “Prompt Engineer” was considered a highly paid profession of the future. That expectation hasn’t quite panned out. Newer models forgive sloppy phrasing better than earlier ones did. Still, careful prompting remains a basic skill, similar to writing a good Google search query.

What belongs in a good prompt

It has proven useful to name four things: the task, the context, the desired format, and the role or tone. So instead of “Explain inflation to me,” try: “Explain inflation in five sentences for a tenth grader, with an example from the supermarket.” The model then has hardly any room left to produce something unsuitable. It’s the difference between a vague shout and a clear assignment.

A second widespread technique is called few-shot prompting. Here, you give the model two or three finished examples before posing the actual task. Someone wanting to automatically sort customer emails into categories, for instance, first shows three correctly categorized emails. The model recognizes the pattern and applies it. Examples often work better than any explanation.

For calculation and logic tasks, it helps to request intermediate steps. The instruction “Think step by step” measurably lowers the error rate. The reason: the model generates text word by word and can build on its own intermediate results. Modern reasoning models now do this on their own, by thinking internally before giving an answer. For them, elaborate prompt tricks are less necessary.

Prompts in apps, image generators, and job postings

You most obviously encounter prompts in the chat window. But they’re also hidden in tools that don’t show any input field at all. When a translation feature in a browser or a summary function in an email app runs, the program sends a pre-built prompt to a model in the background. The user only sees the result.

With image generators like Midjourney, prompting is especially visible. There you describe an image in words, often including details about style, lighting, and perspective. There are entire collections of successful image prompts that users trade among themselves. This shows how strongly results can depend on individual words.

In the news, prompts mostly come up in connection with security problems. In so-called prompt injection, someone smuggles instructions into a webpage or document that the AI later reads. The model can’t reliably distinguish between a hidden command and a genuine user request. For companies, this is one of the hardest unsolved problems in deploying AI assistants.