There's clearly room for improvement there, you're right, but I'd also say these kinds of tasks are severely underconstrained, and if you have a certain image of how it should look in mind, you should give it a lot of references, especially if UI is complex. Screenshots, figma, sketches, references to existing design, etc. And even then there's probably going to be a few iterations.
I've had a pretty good experience tampering with existing UI without taking time of the frontend team, but depending on the scale and your abilities, some things may indeed be faster manually still
On one hand they can't see. Haven't seen a "give me an svg unicorn" test since a while, but I guess their performance hasn't changed a lot. If you give broad prompts, then that will result usually in some garbage.
As an anecdata, a few weeks ago I was creating some UI with libSDL for an embedded project with Claude, and I was quite impressed with the result. Granted, I specified everything I could think about: I instructed it page by page, iterating button by button and label by label.
1. Did not use the UI framework currently being used in the web app;
2. Did not see the pattern, I always use header, but it skipped;
3. Wrong placement of elements, e.g. one element should have been on the left pane, instead it put it on the top right and so on.
I agree that on pure code they are OK, but I thought maybe I am missing something?
That is incorrect for most modern agents, which support multimodal inputs and are trained as such as it is an inherent requirement for things such as computer vision.
There's a reason OpenAI didn't release Codex-only variants of GPT 5.4 and up.
There's clearly room for improvement there, you're right, but I'd also say these kinds of tasks are severely underconstrained, and if you have a certain image of how it should look in mind, you should give it a lot of references, especially if UI is complex. Screenshots, figma, sketches, references to existing design, etc. And even then there's probably going to be a few iterations.
I've had a pretty good experience tampering with existing UI without taking time of the frontend team, but depending on the scale and your abilities, some things may indeed be faster manually still
On one hand they can't see. Haven't seen a "give me an svg unicorn" test since a while, but I guess their performance hasn't changed a lot. If you give broad prompts, then that will result usually in some garbage.
As an anecdata, a few weeks ago I was creating some UI with libSDL for an embedded project with Claude, and I was quite impressed with the result. Granted, I specified everything I could think about: I instructed it page by page, iterating button by button and label by label.
In this case, could coding it manually be faster?:)
what, exactly, were the failures?
coding assistants are not visual-first, they are code-first. so it makes sense that they excel more at coding.
1. Did not use the UI framework currently being used in the web app; 2. Did not see the pattern, I always use header, but it skipped; 3. Wrong placement of elements, e.g. one element should have been on the left pane, instead it put it on the top right and so on.
I agree that on pure code they are OK, but I thought maybe I am missing something?
That is incorrect for most modern agents, which support multimodal inputs and are trained as such as it is an inherent requirement for things such as computer vision.
There's a reason OpenAI didn't release Codex-only variants of GPT 5.4 and up.
[flagged]