What is a System Prompt?

While a user message is a specific request, the system prompt sets the persistent context the model operates under for every turn of the conversation — who it is acting as, what it should never do, how it should format responses, and what tools or data sources it has access to. Because it's processed before user input, a well-designed system prompt shapes the model's behavior more durably than any single instruction buried mid-conversation, and models are generally trained to weight system-level instructions with higher priority than conflicting user requests. Writing an effective system prompt is a distinct skill from writing a single good query: it needs to anticipate the range of situations the model will face across an entire session, specify edge-case behavior (what to do when a tool fails, when a request is ambiguous, when to escalate to a human) and avoid being so long or contradictory that the model starts ignoring parts of it. For agentic systems specifically, the system prompt often also defines the tool-use contract — what tools exist, when to use them, and how to interpret their outputs.