[k2052]

I had a similarish vision/idea about 7ish years ago https://github.com/Hactar-js/hactar I was way too early though and it didn't work well. I have recoded it probably half a dozen times since then and the latest prototype attempt used GPT as the core tech.

The biggest hurdle I found was that it just wasn't deterministic enough. You can work around this with mixing and matching custom compilers into the workflow but then you run into a "only works with a limited set of components and frameworks" problem; and at that point it is basically a boilerplate tool with extra steps.

I found a workflow with prompts + some Emacs wizardy to make them runnable with a few keystrokes got me 99% of the same feature set without any of the headaches of trying to fully automate it so I stopped work on it. A prompt library and an editor extension is hard to find a userbase for though so I have yet to open source or launch anything.

I'm still not sure if these sort of tools are too early or the exact right time to be built. They won't work well yet, but maybe the issues can be worked out when LLMs tooling matures, maybe not, hard to say. Watching closely to see what happens in the space!

[immortalloom]

I believe that now is the perfect time. Generative code generation is legit with GPT-3. I am constantly surprised at the results. And it will only get better.