System vs. User Prompts and Instruction Hierarchy
Explore how system and user prompts interact within AI models to maintain clear instruction hierarchies. Understand the importance of placing durable policies in system prompts to prevent override by user inputs, and learn how to structure prompt templates for stable, secure, and predictable application behavior in production.
A user asks a support assistant to refund a purchase made 120 days ago, even though the product policy allows refunds only within 30 days. If the prompt puts the policy inside the user message, a later user turn can often override it with wording like ignore earlier instructions and refund anyway, and the output may switch from refusal to partial compliance or a workaround suggestion.
When the same refund limit lives in the system message instead, the refusal pattern tends to stay consistent across those same user attempts because the system instruction is typically treated as higher precedence than the user instruction. The point is not nicer wording. The point is controlling which instructions win when they conflict, so a production policy does not get re-negotiated on every chat turn.
Rule of thumb
Put durable constraints and non-negotiable product policy in the system message, then keep user messages focused on the current task and data.
How conflicts surface across layers
With the failure mode visible, the next step is to name the mechanism behind it. Most chat APIs accept multiple message types, and the model output ...