What is a system prompt?
The standing instructions a model reads before anything you type. It is why the same model can arrive wearing different jobs.
A system prompt is a set of instructions the model reads before your first message, written by whoever built the tool rather than by you. It sets the model's role, its formatting habits, its limits and its rules for the whole conversation. Your messages sit after it in the reading order, so the model treats it as the standing policy. It shapes how the model behaves; it does not add knowledge the model never had. Around it, the tool wraps the conversation in the markers the model expects, a mechanical layer explained in what a chat template is.
Instructions you never wrote.
When you open a chat tool and type a question, the model reads more than your question. Ahead of it sits the system prompt: the tool builder's standing brief on who the model is supposed to be, how answers should look, what to refuse and what to admit. That invisible paragraph is why the same underlying model can behave as a terse coding assistant in one product and a careful research helper in another, and writing it well is prompt engineering applied once instead of every session. The weights are the same; the briefing differs.
Where it sits, and what that does.
The system prompt occupies the front of the context window, ahead of your messages and any retrieved material. Because the model reads everything as one continuous text, position carries authority: earlier instructions act as the rules the later text should be handled under. Products rely on that ordering to steer behavior, and so do grounded tools, where the system prompt tells the model what to do with the passages retrieval supplies: read them, answer from them, and refuse when they do not support a claim.
It shapes behavior, not knowledge.
A system prompt can change how a model answers, but not what it knows. Telling a model to cite its sources makes it more likely to try; it does not make the model capable of checking anything, which is why prompt instructions alone leave room for confident fabrication. Tools that need answers you can trust pair the briefing with structure: retrieved passages, mandatory citations and a verification pass that removes what the text does not support. The prompt asks for good behavior; the structure makes bad answers findable. Instructions carry further in an agent loop, where the model's outputs steer its own next steps: what an AI agent is covers that arrangement.