System Prompt
A high-priority instruction that sets a model's role, behavior, or constraints for an interaction.
What stays the same for two gallery visitors?
A gallery-guide application supplies a standing instruction before it passes along each visitor’s question.
- Application instruction
- “Answer from the supplied exhibition notices. If a detail is missing, say which detail needs confirmation.” The application sets this instruction across visits.
- Current question
- One visitor asks for closing time. Another asks where to leave a coat. Each request and its relevant notice are supplied for that visit.
- Result to check
- If the current notice gives closing time but omits cloakroom details, the two answers should reflect that difference.
A behavioral instruction and a permission do different jobs
The application chooses available tools and enforces access to its records. A system instruction can guide how the assistant describes information or handles uncertainty. The exact roles, priorities and supported message formats depend on the model and interface.
Limits of a System Prompt
A system prompt is influential, but it is not invincible. Poor tool permissions, weak application logic, or adversarial input can still cause failures. That is why system prompts work best when combined with guardrails, validation, logging, and permission checks outside the model itself.
In other words, the system prompt is part of the governance layer, not the whole governance layer. It tells the model how to behave, but the surrounding system must still enforce what the model is actually allowed to do.
Follow the idea
Prompt · A prompt supplies the task for an individual interaction.
Guardrails can validate an answer or constrain actions at the application level.
Related concepts
Source · Google’s guide to model-specific roles and prompt structure
Return to System Prompt in the A–Z glossary