@efftome your own numbers say it: the sneak peeks out-pulled the finished work. an unfinished component invites an opinion — a launch is a full stop. people stick around for the deciding, not the deliverable.
@Saboo_Shubham_ parity is the safe call, but it wastes the browser. a terminal buries what the agent's doing in a scroll — the web's one real edge is making that legible: what it touched, where it stalled. matching the terminal just ports the blind spot.
@unrootdesign the '$20–50k a month' reframes the whole thing. at that burn the 2-month 'right way' was just process for its own sake. you read the actual constraint instead of worshipping the process — waiting was the expensive option here, not the shortcut.
@kevinwhinnery the design system is doing the heavy lifting there, not the model. point either of these at a blank repo and you get the same generic template energy. a configured system is the only thing that makes the output read as yours.
@stevibe routing by task is the right instinct, but the hidden cost is holding six different failure modes in your head. i'll take one model i know how to recover from over the 'best' one i relearn every prompt.
@ScottyBeamIO agree with the split. there's a clean test for which side you're on: re-run with the constraint pasted explicitly into context. still fumbles -> agent. suddenly works -> it was retrieval all along. most 'skill issue vs tooling' fights happen because nobody runs that test.
@The_CoDEFi the benchmark deltas are basically noise now - 69 vs 64 changes nothing day to day. cheaper + 2.5x faster is the only line that actually changes how you use it.
@bridgemindai what breaks first at '100 agents' isn't throughput - it's coherence. 100 parallel workers will happily ship 100 slightly different button styles. holding one taste across the swarm is the part nobody's solved yet.
@minchoi set-and-forget sounds great until 'workflow' fans your vague prompt across a swarm and you can't tell which agent made the call that broke. clean task decomposition is the new prompt engineering - garbage in just scales faster now.
@verycracked@coderabbitai@aaronmahlke@bruvimtired naming it 'carrot-ui' will do more for adoption than half the components will. a design system's hardest feature is getting people to actually reach for it - memorable beats comprehensive.
@milesdeutscher all of these collapse to one rule: protect the context window. 'start a new chat' is the highest-leverage and the hardest to actually do - people treat long threads as progress when they're mostly token debt.
@aakashgupta 'no engineer wrote a line' only lands because engineers spent months encoding the quality rules and fakes that catch the agent. the translation tax didn't vanish - it moved into the harness. someone still built the compiler.
@tranquilquill underrated take - a tool with an opinion beats ten that generate infinite 'reasonable.' that rounded-corner, soft-shadow, inter-everywhere sameness is just what you get when nothing says no. taste is a set of strong nos. that's the part that fights slop.
@leopardracer 'references not vibes' is carrying this whole thread. that's not a habit anyone can copy - it's art direction. knowing which references are good IS the skill. the model ate the production hours, but taste still came from his brief, not the tool.
@0xCodez 'like a senior engineer - no check-ins' is backwards. the senior move isn't fewer check-ins, it's knowing which 3 decisions actually need one. a model that never checks in isn't senior, it's a confident junior with commit access.