Вайбкодинг до продакшена: приложение в Google Play, где 99,9% кода написал AI

A developer has published a native Android app, YarrowAI, on Google Play — 99.9% of its source code was written not by a human, but by a neural network. This isn't a weekend toy project: the app is written in Kotlin, contains more than 50 screens, complex animation, multi-language support, and an extensible architecture.

What app did AI write

YarrowAI is a full-fledged native Android app published on Google Play in July 2026. According to the author of the case study, he manually wrote only 0.1% of the code — AI generated everything else. This isn't a prototype or a demo, but a product that has passed the store's moderation and is available to users.

The author separately emphasizes the native aspect: the app is written in Kotlin rather than assembled as a cross-platform wrapper around a web page. This is an important detail — native code across 50+ screens with complex animation is traditionally considered the case where AI generation breaks down most often.

• 99.9% of the source code was written by AI, not by a human
• The app is native, in Kotlin — not a cross-platform wrapper
• More than 50 screens with complex animation
• Support for multiple languages and an extensible architecture
• Published on Google Play under the name YarrowAI (com.yarrowai.yarrowapp)

"This isn't just a 'weekend hack job,'" the author writes in the article on

Habr.

Why AI managed to handle this volume

AI managed to handle 50+ screens because the author prepared an extremely detailed specification in advance, rather than asking the model to "make an app." The level of detail turned code generation from a creative task into a mechanical one.

The preparation involved two key layers. First, a designer drew every screen of the app — the model had an exact visual reference, not an abstract task described in words. Second, for every button the author specified exactly which API needed to be called. With this level of detail, AI doesn't invent the architecture — it fills in a pre-defined framework, and the risk of hallucinations and rework drops sharply.

The case reflects the main shift in "vibe coding": the bottleneck moves away from the code itself toward the quality and completeness of the technical specification. The more detailed the description of every screen and every action, the fewer edits are needed after generation. In essence, the human's role shifts from someone who types code to someone who designs the system and formulates requirements so precisely that the model has no room for guesswork.

Was it worth it

The author directly raises the question of whether the method pays off and suggests looking at the numbers rather than the impressions. Compiling a detailed specification for 50+ screens, with every screen drawn out, is a separate, substantial piece of work that falls entirely on the human and precedes any generation.

"Was it worth it?

Let's look at the numbers," the author writes in the article on Habr.

The main takeaway of the case isn't that AI "replaced" the programmer, but that with strict engineering discipline, AI generation scales up to a real product in the store — not just a single-screen weekend demo.

What this means

The YarrowAI case shows that "vibe coding" has grown up to production use, but not on its own: behind 99.9% AI-written code lies 100% human work on the specification, screen design, and API description. AI takes on the routine of writing code, while the value shifts to whoever can define the task clearly and completely.

Frequently Asked Questions

How many screens does the YarrowAI app have?

More than 50 screens, with complex animation and support for multiple languages.

How was it possible to write almost all the code through AI?

Through an extremely detailed specification: a designer drew every screen, and for every button they specified exactly which API needed to be called.