A writer asked Claude Code, Codex and Antigravity to build an Android app called GearVault from the same prompt, without follow-up revisions. Codex produced working core functionality with some rough interface details, while Antigravity had navigation errors; the writer preferred Claude’s more polished result. This personal evaluation depends on the default models and a single generation, with no controlled repeated runs or cost comparison. Its practical lesson is to test navigation and real user flows rather than treating successful code generation as a finished application.