What is a Prompt?

A prompt can be as simple as a single question or as elaborate as a multi-part instruction set combining a system message, conversation history, retrieved documents, and a specific user request. What the model produces is fundamentally a function of that input — there's no separate "intent" the model reads beyond what's actually present in the prompt, which is why small wording changes can meaningfully shift the output. Production systems typically assemble a prompt from several layered pieces: a system prompt defining the assistant's role and constraints, tool definitions describing what actions are available, retrieved context relevant to the current request, and the user's actual message — all combined into the final input the model actually sees. Understanding a system's behavior often requires looking at this full assembled prompt, not just the visible user-facing message, since most of the behavior-shaping content is invisible to the end user. Because the prompt is the entire interface to the model's behavior, prompt design is a meaningful engineering surface in its own right — see prompt engineering — and it's also the surface that prompt injection attacks target.