Core concept: An MCP server exposes three primitives: tools, resources, and prompts. Prompts are the reusable, user-invoked templates. "Ready-to-use message sequences" is the giveaway.

Why B is correct:

The MCP docs define prompts as "Reusable templates that help structure interactions with language models."
The server-concepts table calls them "Pre-built instruction templates that tell the model to work with specific tools and resources," and marks them User controlled.
"Prompts are user-controlled, requiring explicit invocation rather than automatic triggering." That is exactly "bypass the need for users to write them from scratch."
Clients surface them for easy reuse, commonly as "Slash commands (typing '/' to see available prompts...)". They can take arguments, so formatting instructions ship with the template.

Why the others are wrong:

A: There is no MCP primitive called subagents. Subagents are a Claude Code feature, not part of the protocol.
C: Tools are "Executable functions that AI applications can invoke to perform actions," and they are model-controlled. They run code; they do not hand the user a ready-made message.
D: Resources are "Data sources that provide contextual information," read-only and application-controlled. They supply data, not instructions.

Remember: Tools = the model acts. Resources = the app reads. Prompts = the user invokes a template.

Question 172