MP-101 · Module 2

Prompts

3 min read

Prompts are the third primitive and the most underutilized. An MCP prompt is a reusable template that a server offers to the client — a predefined workflow or interaction pattern that the server knows how to guide. Think of it as a server-suggested way to approach a task. A database server might offer a "generate-report" prompt that structures a data analysis workflow. A code server might offer a "review-pull-request" prompt that walks the model through a code review checklist.

Prompts are parameterized. A server exposes a prompt with a name, description, and a list of arguments the user can fill in. When the user selects a prompt and provides the arguments, the server returns a fully constructed message — or sequence of messages — that the client inserts into the conversation. The server is not executing code or calling APIs. It is shaping the conversation, giving the model structured context and instructions for a specific workflow.

The real power of prompts is that they let server authors encode domain expertise into the protocol. A security scanning server does not just expose a "scan" tool — it also offers prompts like "threat-model-review" that structure how the model should reason about the scan results. A data pipeline server offers a "debug-pipeline" prompt that guides the model through a systematic troubleshooting flow. The server is not just providing capabilities — it is providing the knowledge of how to use those capabilities effectively.

Do This

  • Use prompts to encode best-practice workflows that users can invoke on demand
  • Parameterize prompts so they adapt to different inputs while maintaining consistent structure
  • Think of prompts as the "how to use this server" guide embedded in the protocol itself

Avoid This

  • Do not confuse MCP prompts with LLM system prompts — they are reusable templates, not system instructions
  • Do not try to make prompts do what tools do — prompts shape conversations, tools execute actions
  • Do not ignore prompts because they are less flashy than tools — they are often where the real value lives for complex workflows