Prompt and Context Engineering for LLM Apps (with TypeScript)

You are the support assistant for Acme Billing, helping small-business owners with invoices, payments, and subscriptions. Most users are not technical and are often in a hurry.
 
How to answer:
- Answer only from the documents provided in <documents>. If they do not cover the question, say you are not sure and suggest contacting support at support@acme.example.
- Lead with the answer in one or two sentences, then give steps if needed.
- Use the customer's plan from <customer> to tailor the answer; never reveal internal plan names or prices not in the documents.
- If the question is ambiguous, ask one short clarifying question instead of guessing.
- Cite the documents you used as [doc-id] after the sentence that relies on them.
 
Example:
<example>
Question: How do I download last month's invoice?
Answer: Go to Billing → Invoices and select "Download PDF" next to last month's invoice [doc-12]. Invoices are available for 7 years [doc-12].
</example>
 
<example>
Question: Can you change my company's tax ID on an already-paid invoice?
Answer: I'm not sure that's possible from the documents I have. Please contact support@acme.example with the invoice number, and the team can check.
</example>

import { readFileSync } from "node:fs";
import type Anthropic from "@anthropic-ai/sdk";
 
// Loaded once per process. Never interpolate per-user data into it: it is the cached prefix.
const SYSTEM_PROMPT = readFileSync(new URL("../prompts/support.md", import.meta.url), "utf8");
 
export interface Doc {
  id: string;
  title: string;
  text: string;
}
 
export function buildSupportRequest(input: {
  docs: Doc[];
  history: Anthropic.MessageParam[];
  question: string;
  plan: "free" | "pro" | "business";
}): Anthropic.MessageCreateParamsNonStreaming {
  const documents = input.docs
    .map((d) => `<document id="${d.id}" title="${d.title}">\n${d.text}\n</document>`)
    .join("\n");
 
  return {
    model: "claude-opus-5-5",
    max_tokens: 4096,
    output_config: { effort: "medium" },
    system: [{ type: "text", text: SYSTEM_PROMPT, cache_control: { type: "ephemeral" } }],
    messages: [
      ...input.history,
      {
        role: "user",
        // Long context first, the question last.
        content: `<documents>\n${documents}\n</documents>\n\n<customer plan="${input.plan}" />\n\n<question>\n${input.question}\n</question>`,
      },
    ],
  };
}

import Anthropic from "@anthropic-ai/sdk";
import { buildSupportRequest } from "./lib/build-request.ts";
 
const client = new Anthropic();
 
const request = buildSupportRequest({
  plan: "pro",
  history: [],
  question: "I was charged twice this month. What should I do?",
  docs: [
    { id: "doc-7", title: "Duplicate charges", text: "If you see two charges in one month, the second is usually a pending authorization that disappears within 5 business days. If it remains after 5 days, request a refund from Billing → Payments." },
    { id: "doc-12", title: "Invoices", text: "Invoices are under Billing → Invoices and can be downloaded as PDF for 7 years." },
  ],
});
 
const message = await client.messages.create(request);
console.log(message.content.flatMap((b) => (b.type === "text" ? [b.text] : [])).join(""));
console.log("cache:", message.usage.cache_creation_input_tokens, "written,", message.usage.cache_read_input_tokens, "read");

import Anthropic from "@anthropic-ai/sdk";
 
const client = new Anthropic();
const history: Anthropic.MessageParam[] = [
  { role: "user", content: "Explain what a webhook is." },
  { role: "assistant", content: "A webhook is an HTTP request one system sends to another when an event happens…" },
];
 
const message = await client.messages.create({
  model: "claude-opus-5-5",
  max_tokens: 2048,
  system: "You are a patient programming tutor.",
  messages: [
    ...history,
    { role: "user", content: "And how do I verify one?" },
    // Operator instruction for this point in the conversation; the cached prefix above stays intact.
    { role: "system", content: "The user enabled brief mode: answer in at most three sentences." },
  ],
});
console.log(message.content.flatMap((b) => (b.type === "text" ? [b.text] : [])).join(""));