Chat apps hide the machinery. Behind every answer is an HTTP request: a model name, your text, a limit on how long the reply may be, and a secret key that identifies who pays. This free course gets you to make that request yourself, on a job that looks like real work. The course costs nothing; the exercises make real calls on your own Anthropic or OpenAI developer account, which the provider bills per use (well under a dollar for all of them on a small model), and the provider may ask for a payment method or prepaid credit first.
The course follows one composite company, Larkspur Hollow Residential, an Ohio apartment operator whose coordinator spends about 38 hours a month sorting free-text maintenance requests. In three short acts you weigh a script against a form change, a chat app and a vendor add-on, set up a company API key with a spend limit and handle a leaked key; classify a real work order from the terminal with curl, then forty of them with a supplied Python script, and compare the results with a human's labels; and then meet what goes wrong in a two-week shadow run (truncated replies, rate limits, and a 2 a.m. gas-leak report labeled urgent instead of emergency that a 90% agreement score never warned about), measure the rare costly case on its own, fix it with a safety rule ahead of the model and a human decision (and find that the safeguard costs far more than the model), estimate the monthly cost on two model tiers, and make the go or no-go call. Examples use the Anthropic Messages API and the OpenAI Responses API side by side, as documented in October 2026.
Three short Decision Labs let you make the calls under pressure, and a final build, self-assessed against a rubric, has you triage a results file and write a one-page go/no-go note you can reuse in your own job. Plan on about an hour of reading (each lesson states its time) and a quarter of an hour for the labs; the hands-on steps add about 30 minutes and the optional final build about 20 more, so doing everything takes closer to two hours. Managers can follow a no-code path through the same decisions: no exam question requires running code.