I do this for custom software I build for clients, mostly on the backend though[1]. Started it two years ago, have rapid velocity and, without AI generally take 30 to 60 minutes to implement a new feature that touches backend persistence and front end UI.
‐----------
[1] I did a show HN on a tiny front end component that gets the most use, recently, I think. ISTR most people in the comments said they they preferred React, a few said it was not novel enough to be a paper (they are correct. It isn't). You can see it here: https://github.com/lelanthran/ZjsComponent/
I'm not really in the mood to open up the whole stack I use yet, I still make money bu having a low effort way to build web apps, and due to the way it works it requires a fraction of the tokens required when actually using AI assist (output for the last time I implemented a backend call from scratch was 10 lines, most of which were comments).
As humans augmented with agents write more code, solutions that require less context shifting to get stuff done will win.
A common web stack may include API handlers, OpenAPI spec, generated TypeScript definitions, generated TypeScript client, React logic and effects code, TSX code, HTML, and CSS.
This generally needs filesystem watchers, code generators, transpilers, compilers to get stuff done.
Something that can go from a backend handlers straight to terse markup for reactive UI would be a massive simplification to all this, and a big productivity boost.