Coerce to Interface

Context

You're aiming for either a standardized way of doing something or a verified-to-work approach.

Pattern

Design MCP tool interfaces that enforce structure through their API definition. Required fields, enums, and typed parameters become constraints the agent cannot bypass.

The interface itself documents usage - well-named fields and clear types reduce the need for separate explanations.

This shifts enforcement from instructions ("please include a title") to mechanism (the tool requires a title parameter).

Example

Learning MCP with constrained interface:

{
  name: "add_learning",
  parameters: {
    filename: { type: "string", required: true },
    title: { type: "string", required: true },
    topic: { type: "string", required: true },
    oneLiner: { type: "string", required: true },
    context: { type: "string", required: true },
    examples: { type: "string", required: true },
    tags: { type: "array", items: { type: "string" } },
    related: { type: "array", items: { type: "string" } }
  }
}

Every learning has identical anatomy. No variation possible. Results are immediately searchable by topic, filterable by tags, linkable through related fields.