I guess a lot of us are trying this (naturally) as solo devs, where we can take an engineering-first mindset and build a machine or factory that spits out gizmos. I haven't gotten to the finish line, mostly because for me, the holy grail is code confidence via e2e tests that the agent generated (separately, not alongside the implementation).
I talked about a similar, but slightly simpler workflow in my post on "Vibe Specs".
https://lukebechtel.com/blog/vibe-speccing
I use these rules in all my codebases now. They essentially cause the AI to do two things differently:
(1) ask me questions first (2) Create a `spec.md` doc, before writing any code.
Seems not too dissimilar from yours, but I limit it to a single LLM