I also love using open superwhisper with codex.
When I'm attaching files in a new thread I need to type, but when I'm following up, I just love to speak and it's so much faster.
I tried Superwhisper and WhisperFlow as well, but Nvidia has an open source model so I don't think it makes sense to pay for a open source model.
Open superwishper is an open source project. Links in the comments.
Here's how to get pixel perfect screens from Figma with clean maintainable code using codex.
You need to use the atomic design paradigm.
It forces every component to own its logic. Debugging gets a lot easier after that. LLMs tend to create fewer bugs when the repo follows atomic design.
If you repo does not have proper components, you should just use a existing system like shadcn or mantine.
For Figma
Don’t treat the screen as one thing.
Break it into layout components. Then prompt each part separately.
In a mono frontend/backend repo, I get almost pixel perfect UI with barely any bugs. Everything just connects properly.
For bigger, more complex screens
Go one level deeper.
Start with the layout. Break down every component. Each component gets its own thread.
You get way more control this way. Things stop randomly breaking, and when they do break, you know exactly where to look.
Use this with git worktrees, and multiple threads for multiple component.