The bigger picture, if you're explaining Cowork/agents to someone else: there are roughly three layers. The model is the reasoning engine, it reads your request, decides what needs to happen, and handles anything ambiguous or language-shaped. Tools are how it actually touches the world, running shell commands, editing files, reaching your linked computer, calling an API, the "hands." Skills sit on top as reusable playbooks: natural-language instructions (optionally bundling scripts, templates, or reference docs) that get pulled in automatically when a task matches, so a recurring job gets done the same well-thought-out way every time instead of being reinvented, and re-explained by you, from scratch each session.

the skill decides what to do and when, the script decides exactly how, precisely and repeatably.

when a skill says "run this exact command," I follow that very reliably. The real distinction is where the judgment lives. A skill is read and interpreted by me each time it fires, that's exactly what makes it good at the parts of this job that need judgment: deciding whether a new folder means a new account, figuring out an unfamiliar bank's file layout by looking at it, deciding when something's ambiguous enough to stop and ask you rather than guess. A script has no judgment at all, it's the same fixed sequence of operations every time, which is exactly what you want for the parts where judgment is actually a liability: computing a dedupe hash, parsing "10.689,00" as an Argentine number versus a US one, matching a description against a rule. Those need to produce the identical answer on the 500th run as the first, be fast over many rows, and be something you could hand to another engineer to review and trust, none of which is really what an LLM generating token-by-token is built to guarantee, even when it usually gets arithmetic right.