so we're building a design system. 40 something components. designer hands me the figma file with all the specs and im thinking "perfect, ai can handle this" but actually: it couldnt lol started with chatgpt. copy pasted figma specs, described what i wanted. got back some jsx that looked ok until i actually tried to use it. styles were pretty off, had to fix so much stuff that i might as well have written it from scratch. tried cursor next. better for sure, especially when working with existing code. but anything more complex than a button or input field needed a ton of editing. also tried this thing called verdent that someone mentioned on twitter. it plans out the component structure first which was kinda useful for thinking through the props. but the generated code still wasnt production ready. buttons and inputs worked pretty good. ai saved maybe 20-30 mins per component. mostly just tweaking styles and adding proper aria labels. cards and basic layouts were fine too. ai gets flexbox. still had to adjust spacing and responsive stuff tho. dropdowns and modals? nope. ai generates something that compiles but the event handling is always buggy. spent more time debugging than if i just wrote it myself. animations were useless. ai either ignores them or adds some weird css transition. ended up doing all of them manually anyway. accessibility honestly pissed me off the most. ai will slap an aria-label on things but completely miss focus traps, keyboard nav, screen reader stuff. had to go through every single component. the data table was the worst. ai generated like 300 lines for sorting/filtering/pagination. looked legit. ran it and immediately hit performance issues cause it was doing a full re-render on every sort. deleted the whole thing and rewrote it with useMemo in like 2 hours. estimated 3 weeks to do this manually. with ai took about 2.5 weeks. so saved maybe 2-3 days total? not the massive speedup i was hoping for. biggest issue is ai just copies generic patterns from the internet. we have our own conventions for prop naming, composition, theme tokens. ai kept generating material-ui style code and i had to keep correcting it to match our patterns. now i just use ai for the boring boilerplate stuff. simple components sure. complex stuff no way, faster to write it myself. my current setup: chatgpt for quick questions, cursor for in-editor stuff, occasionally verdent when i need to plan out something bigger. spending like $50/month total. worth it i guess but not life changing. is anyone else doing this or am i just bad at prompting