Switched the copy agent to a newer model for cost reasons. Factual accuracy went up, tone went flat. Everything in the system prompt was the same. Which parts of a voice guide survive a model change in your experience, and which have to be rewritten?
The vocabulary list survives. The adjectives-to-avoid list survives. The examples do not: a new model reads the examples differently and needs new ones that show the same principle.
This is exactly what the evaluation set is for. Run the voice rubric on the new model before switching and you see the flatness as a number instead of as a complaint.
We keep two models in the pipeline for a week after a switch and diff the outputs. The diff is the fastest tone report you will ever read.
Closing the loop: three new examples, same vocabulary list, rhythm rules rewritten as explicit sentence-length guidance. Voice rubric back to where it was. Senna's pattern note was the shortcut.
Happy it landed. For anyone else switching: diff the two models on the same briefs before you touch the prompt. Half the fixes are obvious once you see them side by side.