The system prompt is the standing instruction to a model. It typically says who the model is, what it should and should not do, how to format answers, and which tools it may use. The user's messages then arrive on top of it.
It is where much of an assistant's or agent's behavior is set, which makes it a product decision, not just a technical one. Keep it specific and testable: vague rules get ignored, and long contradictory ones confuse the model. Treat it like code: version it, test changes against real examples and review what changed when behavior shifts.
Do not assume it is secret or tamper-proof. Users can sometimes coax a model into revealing or bending its instructions, so never put credentials in it or rely on it as the only safety control.
Related: Prompt Engineering, AI Agent, AI Assistant, Context Window.