Definition
A system prompt supplies instructions or context separately from ordinary user input to establish how a model should behave for a request or conversation. It commonly defines the assistant’s task, response style, constraints, available context, and rules for interacting with users or tools.
The exact name and priority behavior depend on the model and API. Some interfaces use system or developer messages. Others accept a separate instruction parameter or use a model-specific prompt format.
Simple example
A support application sends a system prompt that tells the model to answer from supplied policy excerpts, state when evidence is missing, and return a structured response with an answer and cited source identifiers. The user’s question and retrieved policy text are added in separate parts of the request.
The application checks that each cited identifier exists and that the current user may access the corresponding policy text.
Why it matters
System prompts give applications a consistent place to define task behavior without repeating those instructions in user-visible copy. That helps keep responses consistent. The prompt can also set terminology and say how to handle missing information.
System prompts use context-window capacity. Version and evaluate them like other runtime configuration because a change can affect many requests.
One important nuance
A system prompt can guide behavior, but it cannot enforce a security boundary. Conflicting or injected input can still steer the model, and the model may reveal parts of its context. Authorization, data isolation, tool permissions, output validation, and approval for consequential actions must be enforced by trusted application or platform mechanisms, even when the system prompt forbids unsafe behavior.