A skill is a reusable recipe that teaches an AI agent how to do one specific task: packaged instructions and, sometimes, supporting files or scripts. The agent does not read it all the time: it sees its name and description, and loads it when the task fits. Having it installed does not mean it gets used.
A question from the app. Pick an answer.
How do you know if the agent used the skill?
Being installed only says it was available. It does not prove it was opened.
Many tools show when they load a skill. And the work should follow its steps.
Sounding sure proves nothing. The proof is in what it did.
The instructions live in a SKILL.md file, with a name and a description. Some skills add supporting files, like a template or an example, and scripts for the parts that must come out exactly the same every time. Where they are installed and how they activate depends on the tool.
A skill is not the same as MCP. MCP connects the agent to outside tools and data; a skill teaches it how to do something. It is not your project's rules file either: those rules guide all the work, and the skill is for one specific task.
Loading every recipe for every task would be a lot of text and would fill the context window for nothing. So the agent first sees a summary, the name and description, and only loads SKILL.md when it thinks the task calls for it. It checks the supporting files when they are needed.
An example: you build a skill to review copy in plain US English for a store that sells across several states. If the description only says "helps with text", the agent may skip it. If it says "reviews customer facing copy in US English: use it before publishing a product page", it is more likely to pick it right there. If it is too broad, it can show up where it does not belong.
A skill being installed only means it is available. If you ask "fix this bug" and you have a skill for writing tests, the agent may not open it. If you want it used no matter what, ask for it by name in your prompt.
And then check. Many tools show when they load a skill, and the work should follow its steps. An answer that sounds confident proves nothing: the proof is in what the agent did.
"If I installed it, it will use it." Not always: it depends on the task and the description. "With a good description, it is certain to use it." It helps, it does not guarantee. "A skill connects the agent to my data." That is closer to MCP: the skill is the recipe, not the connection.