Notes
🔍

MCP Prompts

Tools let an agent act and resources let it read. MCP prompts cover a third thing the server knows and the agent does not: the right way to ask. They are named, reusable prompt templates a server publishes, written by whoever understands that system best, and offered to the user as ready-made actions.

How a prompt is defined and used
A prompt is a declared entry in the server's catalogue, not a string buried in the agent:

  • It carries a name, a description and typed arguments — a review-pull-request prompt taking a repository and a number;
  • The client enumerates them with prompts/list and fetches a filled-in one with prompts/get;
  • The server returns a prepared sequence of messages, which can embed resource contents — the diff, the project's style guide — already fetched and formatted;
  • The host surfaces them as slash commands or menu items, so a person selects one instead of composing the request.

The control distinction is deliberate. In the protocol's terms prompts are user-controlled, tools are model-controlled and resources are application-controlled: the model never decides to invoke a prompt.

Why it matters
Expertise about a system belongs with that system. Without prompts, every agent integrating a server reinvents the phrasing that gets good results out of it, and each copy drifts from the others; with them, the server author improves one published template and every client inherits the fix. For the user it removes a guessing game — the shape of a good request is already on the menu.

MCP prompts are named, reusable prompt templates published by an MCP server alongside its tools and resources. Each declares its arguments and returns a prepared message sequence, often with resource content already embedded. Unlike tools, they are user-controlled: a person selects one, typically as a slash command. They let the server author encode the proven way to ask something, instead of every client reinventing the phrasing.

Some requests are worth writing down once and reusing — like the form at a doctor's surgery that already asks the right questions in the right order.

MCP prompts are exactly that, for AI assistants. When a service connects to an assistant, it can bring along a set of ready-made requests: review this document, summarise last week's sales, check this code for mistakes. You pick one from a menu and fill in the blank.

What the assistant then receives is a carefully worded instruction, written by the people who built that service and know how to get the best out of it. It saves you guessing how to phrase things.