Comment by anonyfox

Maybe I am in a minority position here, but despite me vibecoding for many months now (havent written a single line by hand and forced me todo so in the beginning), I still have my IDE open right next to Codex/CC and while the LLM is crucnching along and doing TDD loops I actually read whats created/changed and just sit with it judding if its only right on surface or semantically stupid underneath, essentially realtime-architecting and steering the code agents sometimes even midflight. so I do end up with these 200k+ LoC projects now since typing is lightning fast and 2/3 of my codebase is tests (I insist on regression tests after every steer) but in fact I perfectly know what each piece is doing and where it is, as well as the still not optimal parts and have a mental list for refactoring it later when I have time or a spare parallel agent can do it when feature work isn't crossing the same areas.

so I COULD take over by hand again like I did the decades before just fine, but I refuse to and instead play a codebase like a RTS - lots of stuff happening in parallel but at all times a understanding where is which thing going on and have the next steps in mind (sometimes directly queued as follow up instructions). For me vibecoding is a strict speedboost and literally gamified projects I work on, and the guardrails not only in textfiles but much more in executable code (linters, tests, dependency checks, playwright, ...) as feedback loops agents can spin on on their own made it all click together to the point my main bottleneck is stuff like the Codex app itself using high CPU and memory on my local mac.