Usage & Prompting
Prompt
Everything the model gets to see before it answers: question, instruction, examples, supplied documents. The prompt is not a command but context — the model completes what plausibly follows. That is why even a different sentence structure changes the result, and why it is worth treating prompts like work instructions: precise, with roles, limits and examples.
An example
“Write me something on the quarterly report” yields an arbitrary text. “Summarize the attached quarterly report in five sentences for the supervisory board, name the three largest deviations from plan with figures, and write nothing that is not in the document” yields something usable. The difference is not magic words but four pieces of information: task, audience, length and limit. You would have told a new colleague all four as well.
A closer look
Technically, there is no difference between instruction, example and data: everything ends up as one token sequence in the context. That is why order and position matter. What is at the beginning and at the end is attended to more reliably than what is in the middle.
Why it matters
In operation, prompts become operating resources, and the same requirements then apply to them as to code: they belong in version control, need an owner and must be tested before a change goes live. In most organizations they instead sit in chat histories, notes and heads. That goes unnoticed until someone leaves or a model switch is due — then it is unclear why a phrasing reads the way it does, and nobody dares touch it.
Commonly misunderstood
Because data and instructions sit in the same channel, an inserted document can contain instructions that the model follows. That is not a marginal phenomenon but the basis of prompt injection.
Where this term does its work in the magazine.
1 article uses the term or takes up its principle.