The alternative to skills is a single instruction that grows every time somebody needs something new, until it is four thousand words, contradicts itself in two places and nobody will touch it.
Splitting it into loadable units fixes three things at once. Each is small enough to review, each is loaded only when relevant so the context window is spent on the current job, and each can be changed without risking every other workflow.
The organisational benefit is bigger than the technical one: a skill is a written-down way of doing something, which means the studio’s method stops living in one person’s head.
Why not put everything in one system prompt?
It grows, it contradicts itself, and every edit risks unrelated workflows. Smaller loadable units are reviewable in isolation and spend the context window on the job in hand.
What makes a good skill boundary?
A job somebody could describe in a sentence and would want done the same way every time. If it needs three sentences and a caveat, it is probably two skills.
