A five-day, hands-on studio during SUTD's term break. Twenty-nine undergraduates — 21 from SUTD, 8 on exchange from Tokyo City University — built working AI prototypes and, harder, built the evidence that says whether those prototypes can be trusted.
The course exists because of one sentence:
"The hardest part of actually using AI is knowing whether the thing actually works."
— the line the master class was designed around
Anyone can now generate a working prototype in under an hour. That half is close to free. So the deliverable here is deliberately two-sided: a prototype, plus a structured body of evidence that it works — tests, rubrics, benchmarks, failure modes, and an honest account of what the thing cannot do. That's what separates a deployable product from a vibe-coded demo, and it's the part that doesn't get easier when the models improve.
It is not a lecture series. Contact time across the whole week is 7.5 hours; the rest is build time, by design. Monday and Friday are in person, Wednesday is a required remote checkpoint, and Tuesday and Thursday are optional query clinics. Students need nothing but a paid ChatGPT or Claude account and a willingness to ship something before they know what it should be.
The prototype was never the hard part — by Monday afternoon, most students already had one. What the week actually built, day over day, was a habit of evidence: real users instead of assumptions, agents that test rather than just build, a checkpoint that asked "how do you know," and a final day spent mining the class's own chat logs for what to do differently next time. Twenty-nine students shipped real, working products. The more durable result was the judgment underneath them.
"AI made building cheap. That made purpose, judgment, communication, testing, user observation, and willingness to change direction more important."
— the lesson the class converged on by Day 5, largely on its own