World's Fair 2026
Stop Writing Tone Instructions. Layer Them. - Isadora Martin-Dye, Isadora & Co
Overview
This talk introduces a four-layer architecture for managing AI agent behavior, particularly focusing on brand voice and user interaction. The core thesis is that a single system prompt is insufficient for complex AI applications, especially where nuanced communication is critical. Instead, layering distinct responsibilities—immutable identity, situational mode, example-anchored voice, and post-generation veto—provides a more robust and reliable system. This approach moves beyond simple prompt engineering to a more structured systems engineering methodology.
Who should watch
- AI Engineers building conversational agents
- Product Managers defining AI product behavior
- Developers seeking to improve AI reliability and brand consistency
- Anyone struggling with AI models producing off-brand or factually incorrect responses
Key takeaways
- AI agents are better conceptualized as brilliant interns with high IQ but low EQ, requiring structure and oversight rather than simple programming.
- A four-layer architecture is proposed: Layer 1 (Immutable Identity) sets hard, unchangeable rules; Layer 2 (Situational Mode) adapts to user context and real-time conditions; Layer 3 (Example-Anchored Voice) provides stylistic guidance; and Layer 4 (Post-Generation Veto) acts as a final deterministic check.
- This layered approach prevents common failures where a single prompt tries to handle too many distinct tasks, leading to errors on edge cases or when user context shifts.
- Layer 1 is crucial for preventing catastrophic errors, such as an AI pretending to have a physical presence or, in sensitive applications, providing false hope to families of missing persons.
- Layer 2 allows the AI to adjust its communication based on the user's role or current situation, ensuring appropriate tone and information delivery.
- Layer 3, while common, is insufficient on its own as it only addresses the "happy path" and cannot adapt to novel situations or enforce hard constraints.
- Layer 4, the veto, is deterministic and essential for catching errors that probabilistic methods miss, ensuring factual accuracy and adherence to critical rules before output is delivered.
- A multi-tenant system requires careful handling of identity to prevent accidental data leaks or misrepresentation, where missing identity should cause a crash rather than a silent fallback.
Notable quotes
*I'm not programming a robot, I'm managing a brilliant intern with an incredibly high IQ and a terrible EQ.*
*The first three layers are all instructions. Identity, conditions, and voice. They're all things you tell the model, and the model usually listens. Usually, they are a request. The fourth layer is not a request. It is reading what actually came out and decides whether to allow it to leave your business.*
Unofficial community note. Prefer the recording for nuance.