AI can get the implementation done just fine but in my experience it needs guidance because it can't guess what features and details I want / don't want.
The difference to "before" is that it's now possible at all for me to get personal projects to a working state but in sum they still require a few hours every now and then supervising.
My personal projects matter a lot. For some of them I am trusting my private data in the code, so I need to hold it to a high security and privacy standard. For others, I have hundreds or thousands of users (game mods, helper tools for hobbies, etc.), and I don't want to ship low quality to them.
If it doesn’t matter then do what you want. By definition, it doesn’t matter.
Some of us actually use these tools, so it does matter. You do have to consider things when it matters.
It’s amazing that you can get something usable quickly, but then every time you use the software you’re back in the AI working on your tool.
Using AI to build the tool becomes the project, replacing the project that led you to think about vibecoding a tool.
For instance in the context of this SketchUp clone that might mean using AI to generate an addon for Blender.
For some of my recent projects I'm using the community edition of Odoo to serve as that substrate because it already has an established community of modules to draw from, ten years or so of real world use, and I don't have to burn tokens generating a user account system with permissions or the data models / ORM to interface with them.
This seems to be serving me well, but we'll see. It could be that I end up in the same the tool becoming the project scenario that you describe, just a little bit later than if I had started from a blank slate.