Vibe Coding a Backyard Garden App With Gemini: What Actually Happened
A Verge writer used Google Gemini to build a backyard gardening app with a single prompt. Here's what worked, what broke, and what it means for non-coders.
LUMIEN4 min read
A writer at The Verge gave Google Gemini a single, detailed prompt and came back five minutes later to find a working backyard gardening app sitting in a preview window. There was also a bug. Gemini flagged it, offered a button to fix it, and closed the issue in 233 seconds with an explanation involving race conditions and channel blockages. The author understood none of the technical language but found the whole process exciting. It was their second or third attempt at building an app this way.
What happened
A journalist at The Verge used Google Gemini to build a personal app for tracking and managing their backyard garden. The process started with one long prompt. About five minutes later, a functional app was visible in a preview window.
It did not go perfectly. Gemini surfaced an error message reading “Channel is unrecoverably broken and will be disposed.” Alarming phrasing, but directly below it was a button to fix the problem. The author clicked it. Gemini resolved the issue in 233 seconds, describing the fix in terms of “blockages” and “race conditions.”
The author stated plainly that they did not understand a word of the explanation. They also described the experience as thrilling. This was noted as their second or third attempt at this kind of prompt-to-app workflow, sometimes called vibe coding.
Why it matters
Stories like this one are becoming a reliable beat in tech media, and for good reason. They show where the real floor is for AI-assisted development right now: you can get something working fast, but the process is not clean.
A few things stand out from this particular example:
- Speed is real. A working app preview in under five minutes from a single prompt is not nothing. That would have taken a beginner days, or cost real money to hire out.
- Errors still happen. Gemini produced a bug on its first pass. The one-click fix is convenient, but it also means the model is shipping broken code and patching it reactively.
- The user is still in the loop. Clicking a fix button is a small action, but it is a decision point. The tool needed a human to approve the repair before moving forward.
- Opacity is a real issue. The author had no idea what “race conditions” meant in this context. That is fine for a personal garden app, but it matters more when the output is customer-facing software.
For small business owners who have been curious about building internal tools, simple client portals, or lightweight tracking apps without hiring a developer, this workflow is worth watching. The barrier is lower than it has ever been. The risk is also not zero.
Our take
At Lumien, we have been testing these prompt-to-app tools with clients for a while now. The honest summary: they work best when the scope is narrow and the stakes are low.
A backyard garden tracker is a perfect use case. It is personal, the data is not sensitive, and a bug means your tomato schedule is slightly off, not that a customer’s order is lost. Scale that scenario up to a booking system or an inventory tool, and the calculus changes fast.
The one-click bug fix is the detail that deserves the most scrutiny here. It feels helpful. But it also means you shipped broken code, then approved a fix you could not read or verify. For a personal project, fine. For anything touching real users or real data, you need someone who can actually read what Gemini wrote and confirm the fix did not introduce a second problem.
Vibe coding is a legitimate starting point. It is not a substitute for review. The gap between “it works in the preview” and “it works reliably in production” is exactly where things go wrong, and no prompt closes that gap automatically.
What to do about it
If you want to try this kind of AI-assisted app building, here is a practical approach:
- Start with something low-stakes and personal, exactly as this example shows.
- Write a specific, detailed prompt. Vague instructions produce vague apps.
- When the tool flags a bug and offers a fix, read the explanation even if you do not fully understand it. Copy it and ask a second AI tool (or a developer) whether the fix makes sense.
- Before sharing any AI-built app with real users, have someone with coding experience do a short review pass.
The tools are getting faster. Basic review habits still matter.