The eval suite is the most honest spec a team has. It's the one place where "good" stops being a vibe and becomes a number someone agreed to. So writing it last — after the feature ships, as a QA afterthought — gets the order exactly backwards.
Write it first. The act of defining the examples, the rubric, and the failure cases forces every ambiguous product question into the open before any code exists. More often than not, the feature designs itself once you know what you're grading.
This is really the same claim as memory being a product decision: the measurable surfaces belong to product, not to the step at the end.