What they did
The poster, who says they are not a programmer, used Google's Antigravity editor with the Gemini 4 Argon model on the AI Pro plan to build an Android app. The first prompt was about 1,200 words describing an app that pulls articles from half a dozen news sites and user-chosen RSS feeds, hides items with blacklisted keywords, never shows the same story twice and has a settings panel for the UI. They then sent about a dozen follow-up prompts of 50 to 750 words for an AGENTS.md file, bug fixes, data-loss cases, an app icon and polish, switching from Medium to High reasoning partway and starting a fresh session near the end. The post logs minutes worked, files touched and lines changed for every step.
How it went
The first prompt took 14 minutes and produced 11 files and about 3,600 lines. In total the agent worked about 100 minutes and the weekly usage meter fell from 100% to 12%. When the poster asked why the app was 30MB, the agent found Android Studio had bundled large UI libraries; after swapping to Android's built-in widgets it shrank to 0.7MB. The poster says the app works well; it has not been shared.
Worth knowing
Commenters pointed out that switching reasoning level mid-session likely threw away the cache, and suggested writing a spec and a plan first to spend fewer tokens. Gemini models run here inside Antigravity, Google's agentic editor, which has no separate tag on this site.