Why v0.dev works so beautifully

So, v0.dev released another version, this time with Three.js integration. Some of the examples they have put in there are amazing.

But why does it work so well? I initially thought they had trained a custom model. Then I did a little digging.

First, I found a few open-source alternatives that, while not as good, were still comparable and often created by just one person. They didn't have any custom models; it was purely prompt engineering. To test this, I used Claude Sonnet to see if I could achieve the same quality as v0.dev. To my surprise, it was just as good. Obviously, this was a general model, which meant it was just pure prompt engineering.

Why is it that just a prompt is good enough? I think it's because of the structure embedded in the libraries used in these artifacts. Shadcn UI, Three.js, Chart Library, and Tailwind CSS have implicit knowledge of their respective applications, providing a structured layer that the LLM can leverage.

So, when people say we will no longer be coding and AI will do all the work—maybe!

But what seems clear to me today is that we need more structured libraries and frameworks so that LLMs can really reduce their cognitive load and just glue together these frameworks to give us useful software.

References

https://x.com/v0/status/1829224860158992721