Prompt

A prompt is not just your question. It is the whole payload: the system prompt (Standing instructions sent ahead of every conversation, setting the role, rules and format.), any documents, the conversation so far, the tool definitions, and the thing you just typed. The model sees one undifferentiated sequence of tokens and does not know which part came from whom.

That framing explains most of what makes prompting work. Being specific about the output format helps because the format is part of what is being predicted. Examples help more than adjectives. And anything in the context — including text inside a document you pasted — is read as input, which is why instructions hidden in a retrieved document are a real problem and not a theoretical one.