A practitioner who built AI agents for wedding venues and other sensitive use cases shares a four-layer prompt architecture that goes beyond simple system prompts. Layer 1 is immutable identity rules that cannot be overridden (e.g., never claim to have a physical body, never use words like 'confirmed' in a missing persons tool). Layer 2 is situational mode — real-time context about who the user is and what they're going through. Layer 3 is the example-anchored voice and tone guide. Layer 4 is a post-generation veto — the only deterministic layer — that reads actual output and can reject it before it reaches the user. The key insight: the first three layers are probabilistic instructions, while layer four is a deterministic permission gate. The talk also covers multi-tenant pitfalls, including a white-label identity leak caused by silent defaults, and argues that brand voice failures stem from asking one mechanism to do four fundamentally different jobs.