Higher-priority instructions supplied by an AI application or platform operator to guide model behavior and safety constraints.
A system prompt is a set of instructions supplied by an AI application or platform to guide model behavior. It can define the role, constraints or operating expectations of the assistant. It differs from an ordinary end-user question, although the exact roles and terminology depend on the provider and application.
Instruction priority is important, but a prompt is not a complete security system. Sensitive operations also need appropriate access controls, validation and approval checks. A sentence telling a model not to reveal a secret is not a safe reason to place that secret in a context where it is unnecessary.
For a business building an AI workflow, the system-level instructions should reflect the actual authorized task. They should not claim capabilities the application does not have or tell the model to manufacture answers when evidence is missing. Clear boundaries help make failures easier to identify and handle, but they still need testing.
Takeaway: Use higher-priority instructions to define behavior, and use real application controls to enforce access and consequential-action boundaries. Do not rely on prompt wording alone to protect a business process.
Example
A booking assistant is instructed to explain available appointment types and ask for confirmation before a reservation is submitted. The application separately controls access to the booking system and validates the request. The instruction supports the workflow; it is not the only protection against an incorrect reservation.
What to check
Document which party controls each instruction layer and which actions require approval. Test conflicting or unexpected inputs. Keep confidential configuration and credentials out of ordinary conversational content unless a secure, necessary mechanism specifically requires them.
For more detail, see OpenAI prompt engineering guidance.