Yes I do agree with this. I believe we are shifting from "make the model good" (prompt/context engineering, etc) to "define good for the model" (success criteria/rubrics). Over time I believe this will become increasingly obvious (as long as model capabilities continue to increase).