Should I build a large library of AI 'skills' before moving to agents, and how many skills is too many?

One firm is heavily focused on building out skill libraries internally before layering on agents, rather than jumping straight to agents. Claude has published guidance suggesting each skill should be iterated 10-20 times before production use, which is described as painful, and that after too many iterations/exceptions, output quality can start going backwards rather than improving. That guidance also suggests that once a system has over 50 skills, it can get confused about which skill to summon for a given request. One participant had built 27 skills so far. The general industry trend observed on X/YouTube/Instagram is to build the skill library first, though many people skip straight to agents instead.