It's kind of a tragedy of the commons situation because if you develop a skill, of course you want to tell the agent to "BE SURE TO INVOKE THIS WHEN USER MENTIONS X" or whatever, but then all the skills add the same language and the agent doesn't know what to do.
I built a skill grader to grade Agent skills based on the paper's methodology. Paste Github repo and see score.
My own skill got a C, by the way, so I'm improving it by reducing keywords and description bloat, etc.
I hope we can get to a place where users all reject the bloated skill trend... it just never made sense intuitively to have thousands of words of context loaded to do simple things.. cough G stack...