MCP and AI basics
Prompt (MCP)
A reusable, server-defined instruction template a user can invoke, for example a 'weekly ad review' workflow that the server supplies. Not to be confused with the text a user types into a chat.
In MCP, a prompt is a reusable instruction template defined by the server, not the text a person types into a chat. A server for an ad platform might ship a 'weekly account review' prompt that asks the model to pull the right reports, compare periods and write a summary in a set format. The user invokes it by name, fills in any arguments and the host runs it.
Prompts are useful because they encode the vendor's or your team's knowledge of how a task should be done. An agency can standardise a client audit this way so every account is checked the same way. They also reduce prompt injection risk slightly, because the instructions come from a source you chose rather than from content the model reads along the way.
Support varies by host. Some assistants expose server prompts as slash commands or menu items, others ignore them. When you compare servers, the presence of good prompts is a sign that the vendor thought about real workflows rather than only wrapping an API.