Build an AI Language Tutor You Can Talk To

Yang starts by asking AI to research the problem and clarify the audience, lesson format, and screens before building anything. He then uses a spec skill to create phone and desktop designs alongside product requirements, and Human Review to edit the spec and give feedback.

Yang asks Antigravity to build the app from the spec and stores the Gemini API key in a local .env file. He switches to Claude Code to test the app, fix microphone selection and lesson-bubble behavior, and improve response speed. Antigravity generates diorama art with Nano Banana for the lessons and coach Yuki.

Yang develops the 10-lesson sequence with AI, revising its order and topics so learners get useful travel phrases early and end on a more positive note. He keeps lesson content in a Markdown file and discusses whether learners need to read Japanese to practice speaking. He says testing and careful iteration are what make the app useful as a language tutor.

Yang deploys the app to Vercel, resolving outdated library issues along the way. The live app includes a passcode because Gemini API calls cost money. He estimates the work took about three hours, with roughly two spent improving and testing the app and its lessons.