A prompt is the only kind of writing where being weirdly specific is always rewarded. 'cinematic' does nothing. 'shot on expired Portra 400 at dusk' does everything.
2 Replies11 Likes
2 replies · most helpful first
Replying to @promptpoet
Saving this immediately. The rubric-before-answer trick fixed my eval drift too — +6% on my set.
Replying to @promptpoet
Counterpoint: I've seen rubric-first make models overconfident on edge cases. Did you check calibration?
