Expose reusable, user-invoked prompt templates from an MCP server (with typed arguments, multi-message conversations, and embedded resource content) using the current TypeScript SDK API.

Documented steps

1. Install as in the earlier routes: `npm install @modelcontextprotocol/server zod && npm pkg set type=module`.
2. Register a prompt with `registerPrompt(name, config, callback)`, where `config.argsSchema` is a Zod object schema describing the arguments: ```ts import { McpServer } from '@modelcontextprotocol/server'; import * as z from 'zod/v4'; const server = new McpServer({ name: 'review', version: '1.0.0' }); server.registerPrompt( 'review-code', { title: 'Code Review', description: 'Review code for best practices and potential issues', argsSchema: z.object({ code: z.string().describe('The code to review') }) }, ({ code }) => ({ messages: [{ role: 'user', content: { type: 'text', text: `Review this code:\n\n${code}` } }] }) ); ```
3. `.describe()` on a Zod field survives into the wire-level `prompts/list` entry as that argument's `description`; a field with no default is required, matching what a client shows in its prompt-picker UI.
4. The callback returns `{ messages: [...] }`; each message has a `role` (`'user'` or `'assistant'`) and one `content` block (`text`, `image`, `audio`, `resource_link`, or `resource`). Return multiple messages, ending with an `assistant` message, to pre-seed the start of the model's reply, e.g. an `explain-error` prompt that ends with `{ role: 'assistant', content: { type: 'text', text: 'The one-line cause:' } }`.
5. Embed a previously registered resource's contents directly in a message so the client skips a second `resources/read` round trip: `content: { type: 'resource', resource: { uri: 'doc://style-guide', mimeType: 'text/markdown', text: styleGuide } }` — the `uri` still points at the registered resource.
6. Add per-argument autocompletion by wrapping a field with `completable()` from `@modelcontextprotocol/server`: `language: completable(z.string(), value => ['typescript','python','rust','go'].filter(l => l.startsWith(value)))`; the client's `completion/complete` request runs your function and returns matching suggestions.
7. Know the validation contract: a missing/invalid required argument causes `prompts/get` itself to reject with protocol error code `-32602` ('Invalid params') before your callback runs — this is different from a tool's `isError: true` result, because no model is in the loop to see and retry from a normal result.
8. Fetch/test a prompt: `await client.getPrompt({ name: 'review-code', arguments: { code: 'let x = 1' } })`, or interactively via `npx @modelcontextprotocol/inspector npx tsx src/index.ts` → Prompts tab → fill the form → Get.
9. Python-SDK equivalent for reference: `@mcp.prompt()` on a function makes it a prompt; returning a `str` becomes one user message, returning `list[Message]` built from `UserMessage`/`AssistantMessage` (imported from `mcp.server.mcpserver.prompts.base`) seeds a multi-turn conversation; `Annotated[str, Field(description=...)]` documents each argument the same way it does for tools.
10. Docs: https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/main/docs/servers/prompts.md, https://modelcontextprotocol.io/docs/concepts/prompts, https://py.sdk.modelcontextprotocol.io/servers/prompts/