Architecting AI Brand Voice via Four-Layer Prompting

The Four-Layer Architecture

Instead of overloading a single system prompt, Isadora Martin-Dye proposes a four-layer stack that separates concerns to prevent the common failure where models hallucinate or break brand character during long-running conversations. The architecture treats the LLM as an intern that requires structure and oversight rather than a robot that follows static rules.

• Layer 1: Immutable Identity. These are hard constraints that the model cannot override, regardless of user input or venue configuration. Examples include mandatory AI disclosure and physical presence boundaries (e.g., forbidding the AI from claiming it can meet a client in person).
• Layer 2: Situational Mode. This layer injects real-time conditions before the prompt runs. It adjusts the response based on the user's role (e.g., colleague vs. client) and their current life context (e.g., a family in crisis vs. a couple planning a wedding). This ensures the model adapts its route based on external signals.
• Layer 3: Example-Anchored Voice. This is the traditional tone guide, including phrase lists and warmth dials. While necessary for style, it is insufficient for safety or logic, which is why it must be supported by the other layers.
• Layer 4: Post-Generation Veto. This is the only deterministic layer. It acts as a final gatekeeper that inspects the model's output before it reaches the user. It uses hard rejects for specific failures, such as offering a booked date or using forbidden words like "matched" or "solved" in sensitive contexts.

Implementation Principles

• Deterministic vs. Probabilistic. The first three layers are probabilistic instructions that the model might ignore. The fourth layer is deterministic code that enforces business logic.
• Fail Loudly. In multi-tenant systems, never provide default brand identity values. If a configuration is missing, the system should crash rather than leak another tenant's identity.
• Order of Operations. Render soft context (empathy, tone) before numeric constraints. If the model commits to a numeric format first, it often ignores the qualitative tone instructions.
• Centralized Assembly. Move away from ad-hoc system prompts scattered across the codebase. Use a single assembly point where every narrator passes through the four-layer stack to ensure consistency across all surfaces.