Skip to main content
A prompt owns reusable instruction text. It can include template variables and can be exposed through MCP. Prompts exist because assistants often need named instructions that are not tools. A prompt tells an assistant what to do or how to frame a task. It does not execute code and does not own state.

​
Characteristics

  • Content contains the instruction text.
  • Variables let the caller fill in task-specific values.
  • A stable ID lets MCP clients discover the prompt.
  • The caller decides when the prompt is useful.

​
Boundary

A prompt is read and applied by a caller. An agent or MCP client decides when to use it. A tool owns an action. A resource owns readable data. A prompt owns instructions. This keeps instruction templates separate from executable capabilities.

​
Wrong fit

Do not use a prompt when the project needs to fetch data, mutate state, or call an external system. Use a resource for readable data and a tool for actions. For API details, see veryfront/prompt.