System prompt tone instructions fail in production for the same reason a comment that says "make it work" fails in code: they describe intent without enforcing behaviour. They hold for the first twenty turns and break on turn twenty-one, when the user asks something the examples did not cover.
This talk presents a four-layer prompt architecture built across five production AI products on a shared technical spine, each with distinct voice requirements that had to hold under real load. The immutable identity layer: what the system structurally cannot say, enforced above the model. The situational mode layer: what shifts when user state shifts, triggered at runtime. The example-anchored voice layer: where most teams stop, and why it is not sufficient on its own. The post-generation veto layer: the cheap final pass that catches what the other three missed.
Real prompts shown in full. The multi-tenant pattern that lets a single API spine serve clients with completely different voices without forking the architecture. The failure cases that forced each layer into existence. Useful for any team whose AI agent holds its voice in demo and loses it in production.