80% of my product is AI-generated. Some lessons about LLM-driven building.

When I started building GetStory.io, I couldn't ignore the hype surrounding AI, especially code generation development tools. I used them quite actively. And here are some insights I’ve got from my experience until now:

• The website/landing page is 100% vibecode on the v0 platform with a free account. Except for the deployment - I just did an rsync to my server. The whole process took a little over three hours. This is delightful, of course. BTW, I generated the initial requirements using my own product. It's also funny that some blocks improved simply by repeating the request: "Make it cool! Even cooler, like the coolest startup." After 4-5 iterations, I got what I wanted.
• The product itself, front and back, was made from own boilerplate/template (assembled stack + CI/CD settings). However, 80% of the features, business logic, and UI are generated through LLM.
• However, it's important to note that you still have to deal with architecture, refactoring, and code reviews, as well as the product itself, its business logic, and its UX. In short, all the most interesting and valuable work can (and should!) still be done independently.
• Why is that? Because AI tools quickly become overwhelmed, add solutions that break previous features, misinterpret and fix bugs, and solve similar tasks with different approaches. As a result, chaos in the code base grows like a snowball.
• Nevertheless, LLM replaced a team of 2-3 mid-level developers who could build features in tens of minutes rather than tens of hours. But that's because I understand what's going on. It's not yet clear how nontechnical vibecoders should proceed.
• LLM excels at creating UI components, which is a huge relief because I no longer have to spend hours messing around with Figma and CSS.