What it means
A prompt is everything the model sees before it responds: the question, but also the system instructions, the conversation so far, any documents pasted in, and any examples provided. All of it occupies the context window, and all of it shapes the answer.
The common misconception is that a prompt is a search query. It isn't. A model has no index to look things up in — it continues text. So the prompt is less a request than a setup, and its job is to make the desired continuation the most probable one.
That is why supplying material beats asking harder. A model given the relevant document answers well; the same model asked more forcefully from memory just guesses more confidently.
Why it matters
The prompt is the main control surface most people have. It also carries a security consequence: because instructions and data arrive as one undifferentiated stream, anything you paste in can contain instructions, which is the basis of prompt injection.
In practice
Give context, state the format you want, and provide an example if the format matters. Then treat any external content in the prompt as untrusted input.