If it matters, I'm asking this to gather information to make a decision, I received an offer to work in a company that requires all code to be crafted this way.
I already have a good job and I'm happy with it, so besides a much bigger/trendier name in my CV and more pay there's not a lot in this opportunity for me. I'm also aware that if I were to leave my current position my team could be heavily affected since I usually play a key role in helping us deliver work on time (not that my company would do the same for me, but at the end of the day I don't want to willingly screw my coworkers' chance of success/stability).
As for this new opportunity: What concerns me is that I'm worried about the scale/size of the applications and if current SotA LLMs are capable of building something that's maintainable in the long run. They're a big name company and this work isn't for any of their main products, so I fear that this might just be a pet project for them and if it doesn't work they will scrap everyone involved in the project in a few months.
Sorry for asking this on a public forum, unfortunately I don't have folks in my circle that could help me w/ this question, so all I can say is thanks in advance for your help :bow:!
PS I tried googling and didn't find conclusive results to this question (this is a polarized topic!), it doesn't help that most links I found in popular reddit threads on this subject are dead, the most notable app I found was WithAffluent
Cursor with Gemini 3. I didn't manually write a single line for it (except for the actual wording). Gemini did the React stuff, styling, writing ffmpeg commands for re-encoding the videos to work better with scroll-based scrubbing, splitting them out into AVIF image sequences for mobile, etc. I use Cursor + Codex for the actual game as well but it involves a lot of manual work - even with a really modular system like Unreal it gets convoluted.
I tried doing a spreadsheet application heavily using Sonnet 4 (https://app.embedsheet.ai/) and found it would make a ton of mistakes and massive files that it would then be unable to reason with, I think if I did it again I'd do all the core stuff by hand.
https://tendayweekcalendar.com/
I did no manual coding except small bug fixing.
They're something like electric vehicles - faster, cheaper, safer, with a little asterisk because they're so different and there's plenty of exceptions that people are happy to provide. But basically, if you use them as autocomplete machines, you'll be happy. Design the architecture, don't let it architect you. All the scaling issues is because it gets you to 10k, 100k, 1 mil LOC faster than typing it by hand and you will hit the issues of having to deal with all the LOC.
You absolutely should not use it on things which it's not trained on, like custom SDKs and say, integrating GPT-5. It will hallucinate like hell. It's the equivalent of trying to drive your EV into a river.