The part that makes skills work rather than just being saved prompts is that they live in files you can version. A prompt pasted from a note drifts silently; a markdown file in a repo shows up in a diff when you change it, so when the output changes next month you can see what you did. That is the same reason prompt versioning matters in production LLM systems, just applied to your own workflow. The failure mode to watch as the set grows is overlap - two skills whose descriptions both plausibly match what you asked, and you get the wrong one with no error. Keeping each trigger narrow and mutually exclusive is worth more than making any individual skill smarter.