@mweinbach true i guess it depends what you're looking for. it can match a style and run with it but often the result is uglier than if it had been copying the mockup 1 to 1
@ChrisSeltzer@jxnlco@davis7 unfortunately it does need to be told to use subagents, either in the live prompt or skill. on one hand good because it doesn't chew through quota without user approval but also kind of annoying
i turned the workflow into the `reference-ruler-ui` skill, part of my Subtractive UI plugin here:
https://t.co/drJXNcdI7c
short version: freeze scene, build object graph, extract rulers, screenshot compare, recursively repair.
gpt-5.5 is not bad at UI. vague image-to-code workflows are.
when codex goes from "make a simple app landing page" to deploying on cloudflare pages, wiring a mailerlite waitlist, fixing cors, dns, spf, dkim, dmarc, and email routing ๐คฏ
@doseofpurple_ there used to be full page spreads of the author's other books or publisher's other books where you could even cut out the page and send it in to order them
when codex goes from "make a simple app landing page" to deploying on cloudflare pages, wiring a mailerlite waitlist, fixing cors, dns, spf, dkim, dmarc, and email routing ๐คฏ
@Ferbin08@testingcatalog i haven't run into any issues. because it uses your regular chrome session (and perhaps because it's driven by assistive tech) it can do a surprising amount without being blocked. most sites' bot detection is simply not that advanced.
@argofowl gpt thrives on detailed prompts but it's hard to be comprehensive enough as a human. so briefly describe what you want and have an agent write a detailed prompt, then start a new thread with that prompt.