Last October, I discovered Claude Code and thought: "I'll build a full mobile app solo in weeks." React Native, Supabase, WebSocket, voice pipeline. AI will handle everything.
The first hours were magic. Boilerplate that takes 2 days? Done in 20 minutes. A React Native component with error handling, TypeScript types, responsive styles. Almost perfect on the first try.
Then I started building for real.
The Frankenstein problem
AI generates 400 lines for a screen. It works. Next screen. Works too. You keep going. After 3 weeks: 50 generated files, 15 services, 8 React contexts.
And then you realize something.
Nobody thought about the architecture. Not the AI. Not you either, because you were moving too fast. You have a functional Frankenstein monster. Each piece works individually. Together, it's spaghetti.
I spent more time refactoring AI-generated code than I would have spent writing it myself. That's the sentence nobody says about vibe coding.
What AI actually does well
The repetitive, well-defined stuff. That's where it's a genuine 10x multiplier:
- Boilerplate: component scaffolding, CRUD operations, API endpoints
- SQL migrations: schema changes across 3 database schemas, 30+ tables
- Translations: 1,993 keys across 6 languages (FR, EN, DE, ES, IT, PT)
- Tests: unit tests, edge cases, mocking
- Documentation: README updates, inline docs, API specs
No exaggeration on the 10x. These tasks went from hours to minutes.
What AI misses: decisions
The hard questions. The ones where "it depends" is the honest answer:
- What database architecture keeps 30 Supabase tables maintainable in 6 months?
- How do you structure a voice pipeline (STT, LLM, TTS) so it stays modular?
- Where's the boundary between a WebSocket backend and a React Native frontend?
AI gives you an answer. Often plausible. Rarely the right one. And if you lack the experience to spot the difference, you find out 200 commits later when everything is coupled and nothing is testable.
The balance I found (900+ commits later)
Six months and 900+ commits into building TAMSIV (a voice-powered task manager for Android), here's my workflow:
- Let AI generate the skeleton. First draft, boilerplate, repetitive patterns.
- Review every structural decision myself. Architecture, data flow, service boundaries.
- Refactor early, not late. The moment something feels wrong, fix it. Don't wait for 50 more files to depend on it.
- Never confuse execution speed with thinking speed. Vibe coding accelerates the first. Not the second.
The real takeaway
AI didn't replace the developer. It replaced the parts of the job I didn't enjoy. The rest, the hard choices, the architecture trade-offs, the UX that makes someone come back tomorrow, that's still on you.
Vibe coding is a multiplier, not a replacement. And the multiplier only works if you know what you're multiplying.
TAMSIV is a voice-powered AI task manager for Android. Free on the Play Store. Built solo with React Native, Supabase, and a lot of refactoring.
Top comments (0)