Claude Skills Still Need Baseline Proof

Hot take: skills are useful, but the hype only holds up if they beat a good prompt baseline. Without that comparison, a "successful" skill can just be a more reusable prompt in disguise.

• –Anthropic’s own guidance says to measure performance without the skill first, then compare against it, which is the right bar.
• –Skills are strongest when they encode repeatable workflows, formatting rules, and team conventions you do not want to restate every session.
• –Early benchmark work like SkillsBench suggests curated skills can materially help, while self-generated skills often barely move the needle.
• –For ad hoc CLI work, a strong prompt may already cover most of the value, so the extra authoring overhead is the real tradeoff.
• –The payoff grows when a team shares the same skill, because the process becomes portable across chats, users, and models.