Module 2 · Designing servers agents use well · Lesson 5

Resources, prompts and asking for input

Expose read-only context as resources with cache hints, workflows as prompts the user picks, and ask the user to confirm an action through a multi round-trip request whose state the server signs.

60 min 2 videos 2 exercises 8-question quiz Answers

Setup for the exercises:Install Claude Code Node.js, pnpm and TypeScript MCP servers in Claude Code The Larchmere Desk mock API

By the end of this lesson you can

Choose between a tool, a resource and a prompt by asking who decides when each one is used
Expose resources and resource templates with cache hints, and link to them from tool results
Write a prompt with completable arguments that carries a workflow without carrying untrusted text
Ask the user to confirm an action through a multi round-trip request, and protect the state that request carries
Explain why roots and sampling were deprecated, what replaces them, and which hosts support resources and prompts

A support agent asks the model to reply to a customer. The model drafts the reply and calls reply_to_customer, and Desk emails it at once. Tool annotations told the host this tool reaches the outside world, but annotations are hints, and a recent study of 1,723 host applications found that "only 37.2% gate tool execution behind a blocking approval step" (Majeed et al., 2026). If a confirmation matters, the server has to ask for it. This lesson adds the rest of MCP's server surface to larchmere-desk: resources for context, a prompt for a workflow, and a confirmation that works without the server holding any state.

Resources, prompts and asking for input

Expose read-only context as resources with cache hints, workflows as prompts the user picks, and ask the user to confirm an action through a multi round-trip request whose state the server signs.

Resources, prompts and asking for input

Resources, prompts and asking for input

Sign in to keep reading.

Sign in to keep reading.