What is the difference between a skill, a tool, and a prompt?

These three get confused, so it is worth being precise.

• A one-off prompt is retyped every time and disappears after the turn.
• A tool (see tool use and MCP) gives the agent access to a capability.
• A skill gives the agent the method: the procedural knowledge for using that capability well in your situation.

One comparison puts the distinction sharply: a system prompt is static text loaded once at session start, while a skill is a file-based module the agent discovers, evaluates for relevance, and loads dynamically, including code it can run.^[1] The same guide notes that an integration gives a model access to a tool but not the procedural knowledge of how to use it for your workflow; skills supply that method.^[1] Put concretely: an MCP server might give an agent a tool to query your database, but nothing about that tool tells the agent your team’s naming conventions for a report, which fields matter, or what “done” looks like for this specific kind of request. A skill is where that context lives, written once and reused every time the task comes up again.

The persistence half of that sentence is easy to undervalue until it’s missing. Without a skill, every new conversation re-derives your team’s conventions from scratch — sometimes correctly, sometimes not, and there’s usually no signal at the time which one just happened. Write the convention into a skill once and it loads the same way for every user and every session after, which is what turns a single correct answer into a repeatable one. That gap, between a capability existing once and a capability being reliable on demand, is most of what a skill is actually buying you.