Agent skills rewritten without feedback lose 15 points on SkillsBench
A preprint finds that when AI models repeatedly tidy up expert-written agent instructions with no testing, the instructions get worse. The first knowledge to go is the part the model could not have written itself.

