Peter Steinberger, the OpenClaw steward now at OpenAI working on agents, is highlighting a workflow that pairs ChatGPT's April 2026 image model upgrades with code generation: create precise typography and layouts first via images, then feed those outputs to models like Codex for UI implementation.
This is one of the most powerful ways to leverage ChatGPT Images and somehow almost no one is talking about it
It’s exceptional at typography and layouts
If you think codex sucks at design, try "use imagegen to re-imagine this design and implement that".
Code models often falter on dense layouts and text rendering, but the gpt-image-2 improvements in typography and object placement let developers bypass those limits before the code step.
Steinberger and Lukasz Kaiser have surfaced the combined approach in recent posts, yet adoption numbers, prompt specifics, and how it stacks against alternatives stay unmeasured in available details.
Positive users praised the shared prompt for Codex image generation as a helpful hack that worked like a charm, while negative users called the outputs still horrific or horrible.
No Digg Deeper questions have been answered for this story yet.
@ymir_ke for once I cannot cause it’s an internal tool.
@feifei_qiu @steipete I am using http://quiver.ai for icons - and build a skill so codex/claude can generate svg's via http://quiver.ai
@steipete Peter, you are not in the position to say this when you look at the openclaw app😭
If you think codex sucks at design, try "use imagegen to re-imagine this design and implement that".
@steipete I don't know where this "Codex sucks at design" started (it doesn't) but clearly these are the same people typing prompts like "make me an app in one go" or "no mistakes".
If you're treating your AI like a pet and then evaluating an AI, I got nothing more to say.
@warresnaet Oh man it's already SO MUCH BETTER.
My plan totally worked. There's like 10 people cookin' rn!
@struggler_rick @ymir_ke You don’t need a screenshot of my random internal project to use the image trick for better designs.
@steipete I like to use Claude Code to generate HTML design variants so that I can choose one and give it to Codex to work with. @trq212 taught me the beauty of generating HTML designs in his article:
@steipete Yep gpt image 2 is amazing at mockups
@steipete Can you post a screenshot of this?
Would be cool to see an example, I don’t feel this has been a good approach, but I haven’t tried lately.
@sauna9x @steipete @ymir_ke Lol I think he means the tool he's working on to improve the design is internal and cannot show the screenshot.
This is the imagegen skill https://github.com/openai/codex/blob/main/codex-rs/skills/src/assets/samples/imagegen/SKILL.md
@steipete Brother... Use it
@feifei_qiu @steipete imagegen an icon，then use https://github.com/xzhih/icon-tracer
@steipete How codex can magically make good svg icons ?
@feifei_qiu @steipete Use SF Symbols. https://developer.apple.com/sf-symbols/
@steipete @ymir_ke What's the point of posting this if the general public have no access?
@steipete Here is an example of a project I'm working on now that GPT 5.5 did. It's not perfect, but it is decent.
@steipete Add @pbakaus 's impeccable and you're golden
@steipete @ymir_ke How am I trying it if it's your internal tool?
@steipete Interesting idea. I've just been using "Design this like if Jony Ive and Picasso had a baby" and it's been working pretty good
