Skip to main content

Claude API · hands-on

Learn the Claude API by fixing a support queue that is drowning.

Northwind Outfitters takes 4,100 support tickets a week and sorts them by hand. Two attempts at automating it failed. A child’s injury report sat unrouted for three days because it opened with “probably nothing.” You build the service that fixes that — and every capability you learn is one the fix actually needs.

See it runningStart Lab 1

Or read the scenario first — every design decision in the labs traces back to it.

Start with however long you have

Two of these three need nothing but a browser.

60 seconds

Watch it work on your own words

Walk into the fictional shop, tell its support desk something went wrong, and watch a live model categorise it, rank its urgency, and decide whether a person needs to see it.

Open the storefrontNothing to install. No key.
20 minutes

Take the pieces apart

Break the classifier with a prompt injection. Step an agentic loop one tool call at a time. Find the cache bug that quietly doubles a bill. All of it runs in the page.

Open a playgroundNothing to install. No key.
An afternoon

Build the whole service

From your first API call to an eval harness with an LLM judge, against one coherent codebase you can read end to end. Solutions included for every lab.

Start Lab 1Node 20+, an API key, $2–4 of tokens.

Four routes, four capabilities

One domain. Each route introduces exactly one new idea and builds on the one before it.

POST /v1/triage
Structured outputs

The model's output contract is your type system

POST /v1/resolve
Tool use

Claude queries your systems and shows its work

POST /v1/draft
Streaming

Token-by-token delivery over SSE, with real cost accounting

POST /v1/estimate
Token counting

Know the bill before you pay it

Poke at it before you commit to it

Eight interactive tools, all of them running in the page. Three worth starting with:

The trust boundary

Try to talk the classifier into ignoring its instructions. Some of it works.

Agentic loop stepper

One tool call at a time, with the exact message list at every turn.

Spot the cache bug

One line in the prompt kills every cache hit. Find it before the bill does.

See all eight

The labs

Roughly four hours end to end. Solutions included. You need Node 20 or newer, an API key with billing enabled, and about $2–4 of tokens for the whole sequence.

1Your first call, and reading usage20 min2Structured outputs and schema design35 min3Tool use and the agentic loop45 min4Streaming and SSE30 min5Prompt caching and cost35 min6Evals and LLM-as-judge45 min