Devin builds a speaker name quiz

At AI Engineer World's Fair, Scott Wu demonstrates a small name-learning game made with Devin. The input is a TSV of speaker names and photographs. Each round should display two different faces and one name, allowing the player to choose the corresponding photograph on a phone.

Devin plans the task, creates a project directory, starts a React application and reads the data file. It then builds and deploys an initial version. Wu tries the interface and sees that both photographs still display their names, revealing the answer. The first deployment therefore does not fully meet the intended game behavior.

He asks the agent to hide the labels until after an answer, restyle the replay button and add a streak that resets on a wrong answer. The final deployed demonstration shows a correct choice increasing the streak and a deliberately wrong choice resetting it.

This is a bounded builder demonstration with visible human feedback. Wu's estimate of roughly 170 speakers describes the dataset, not the model's general ability to recognize faces. The practical outcome is an interactive quiz refined through actual play, with the human identifying requirements that were missing or unclear in the first request.

Prompt
Original prompt

Source: ai.engineer · Scott Wu