The research on AI agent skills

Everyone is telling you to install their skill. Almost nobody tells you what one is, or what got caught in January. This is what I found, with every source named and linked so you can check it yourself.

What is inside a skill, the named cases and what each one actually took, what the three published audits measured, what Anthropic says in its own docs, and the prompts I use to read a skill before I run it and to write my own instead.