its certainly benchmarkable.

make a bunch of "build the feature" vs "heres a detailed spec" prompts, and see which one does better.

if you take out the AI boogieman, and instead go with "next token prediction" then "people are doing it wrong" is much more reasonable