▲ Claude Academy
Sign in

Vault note

Track 3 — Building with the API

This is where you stop using Claude and start building with it. You'll learn the Messages API end to end — sending messages, system prompts, multi-turn state, streaming — then the three capabilities that turn a chatbot into an application: tool use (letting Claude call your functions), structured output (getting reliable JSON), and prompt caching + RAG (making it fast, cheap, and grounded in your data).

Audience: developers. · Est. time: ~10–14 hours over a week or two. · Prereq: track-2-effective-prompting and a working dev toolchain (setup-checklist, section c — Node/Python, an ANTHROPIC_API_KEY, git). · Outcome: you can build real applications on the Messages API.

Sequence

The whole track mirrors one deep course — building-with-the-claude-api — broken into four builds. Do them in order; each project layers on the last.

1. Your first API call · ⏱ ~2 h

  • Course: building-with-the-claude-api — modules on the Messages API shape, request/response, system prompts, model selection, parameters (max_tokens, temperature), multi-turn conversations, and streaming.
  • Project: p05-first-api-call — make your first authenticated call, hold a multi-turn conversation in code, and stream a response token by token.

2. Tool use — let Claude act · ⏱ ~3–4 h

  • Course: same course, tool use / function calling modules — defining tools with JSON Schema, the tool-use loop (Claude requests → you execute → you return results → Claude responds), and parallel tool calls.
  • Project: p06-tool-use-weather-agent — build a small agent that calls a real tool (e.g., weather + calendar) in a full request→execute→return loop.

3. Structured output — reliable data · ⏱ ~2–3 h

  • Course: same course, structured output modules — getting dependable JSON, schema-shaped responses, tool-use-as-extraction, and validation.
  • Project: p07-structured-extraction — extract structured records from messy unstructured text into a validated schema.

4. Make it fast, cheap, and grounded · ⏱ ~3–4 h

  • Course: same course, prompt caching, batch, and RAG modules — caching stable context to cut cost/latency, the Batch API for high-volume jobs, and retrieval-augmented generation to ground answers in your own documents.
  • Project: p08-prompt-caching-and-rag — build a tiny RAG pipeline (retrieve → assemble context → answer with citations) and add prompt caching to the stable parts.

You're ready to move on when…

  • You can write a multi-turn, streaming Messages API call from scratch with a system prompt.
  • You can define a tool with JSON Schema and run the full tool-use loop, including returning results back to Claude.
  • You can get validated structured JSON out of Claude reliably.
  • You can explain when prompt caching, the Batch API, and RAG each pay off — and you've built a small RAG flow.
  • You've completed p05, p06, p07, and p08 and they're committed to git.

See also