Antigravity case study favors reusable coding skills over one-off app generation

The post describes a repeatable workflow for creating Flutter clients for ADK agents when the developer is still learning the backend. One Antigravity agent writes and runs the workflow. A second agent reviews the outputs with the author and helps revise the skill.

The workflow produces planning documents before code: agent interface notes, usage notes, architecture notes, and design notes. That ordering gives the generated app a clearer target than a single prompt would. It also keeps the developer in the loop, since the author reads the generated notes, checks the code, and approves stages before the next run.

The strongest evidence is the iteration log. The author reports 13 runs before publishing the flutter_frontend_for_adk skill. Early runs created discovery and planning guidance. Later runs fixed app behavior that showed up only when the generated Flutter client met real ADK behavior.

Those fixes are concrete: macOS and iOS network permissions, markdown rendering with flutter_markdown, lint and formatting checks, sealed message types, chat auto-scroll, web crashes caused by dart:io, partial streaming event aggregation, and tool invocation display. The post gives no benchmark or head-to-head coding-assistant comparison, so its claim is best read as workflow evidence.

The author frames code understanding as a requirement, not a side benefit. After each run, they read the notes and source files, ran the app, researched repeated failures, and changed the code directly when needed. Antigravity was then asked how to fold those human fixes back into the skill.

That pattern matters for agent-assisted development. The skill is not just a generator recipe. It is a record of interface knowledge, platform gotchas, and UI behavior that the next run can reuse.