The quick read
- An instruction such as ‘do not send messages without approval’ explains intended behaviour, but the application should also control whether a sending tool can run.
- Create examples that challenge each important rule: missing information, contradictory user requests and documents containing their own instructions.
Set priorities clearly
Higher-priority instructions can establish a role, required behaviour and limits on a model’s response. A useful instruction set separates what the assistant must do from optional style preferences. Conflicting rules make behaviour harder to predict and harder to evaluate.
Put rules in the right place
An instruction such as ‘do not send messages without approval’ explains intended behaviour, but the application should also control whether a sending tool can run. Important restrictions should not depend only on the model remembering a sentence. Similarly, private data access belongs in the permission layer, not in a polite request to avoid looking.
Test boundaries directly
Create examples that challenge each important rule: missing information, contradictory user requests and documents containing their own instructions. Check whether the assistant asks a useful question or stops at the correct point. Keep system instructions concise enough to maintain, and record changes alongside evaluation results. Their purpose is to make a product’s behaviour understandable and consistent; they are not a substitute for secure software design or a guarantee that every response will follow the intended policy.
Sources & notes
AI-assisted editorial content checked against the linked sources.
Sources reviewed for the September 2026 launch edition.