Stop Writing Tone Instructions. Layer Them. - Isadora Martin-Dye, Isadora & Co

TL;DR

• 1

  Single system prompts fail at turn 21: Your examples cover the happy path, but eventually the model produces something technically correct that your brand would never say.

• 2

  Layer one is immutable identity, not preferences: Rules that cannot be overridden by any configuration, like never claiming physical presence or disclosing AI status upfront.

• 3

  Layer two handles real-time conditions: The same AI speaks differently to couples versus staff, and adjusts tone based on context like a client whose mother is in chemotherapy.

• 4

  Layer four is deterministic permission, not probabilistic instruction: The only layer that reads what actually came out and can reject it, catching failures like offering dates that don't exist.

• 5

  Multi-tenant identity must never have defaults: A missing brand identity should crash the system rather than fall back to another venue's voice, which erodes trust silently.

• 6

  The stakes vary wildly but the architecture holds: A wedding venue risks mild embarrassment; a missing persons tool risks telling someone their child was found when they weren't.