If you use Claude mostly for chat and coding, the easy mistake is judging a skill by prompt length and picking the one that looks smartest on first read. Bad heuristic. Better AI skills are thin on top and thick underneath [C002].
alirezarezvani / claude-skills is the cleanest example I've seen [C001]. Its written standard says SKILL.md should stay under 10KB. That sounds like a minimalism flex. It isn't.
The real move is separation. The top file is the workflow. Knowledge lives in reference docs. Repeatable actions live in scripts. The README describes the stack as SKILL.md + 593 CLI scripts + 711 reference docs, reused across 13 AI coding tools.
Why this matters: one giant prompt is harder to update, harder to reuse, and easier to break. The interesting part is not that the skill looks shorter. It's that the author refused to ship the whole system as one huge prompt.
Boundary: this is about the repo's written guidance, not a live benchmark. And thin on top only works when the support underneath is thick. Otherwise you did not design a better skill. You just hid the mess. Share this with anyone still measuring skill quality by prompt length.