Here's a post copied from the old Rebolforum, which I think may be useful perspective:
2026-02-09 15:36:53
Last week I had another great experience using GPT to add very deeply specified front-end styling to a large project, which had far deeper existing functionality already fully built.
I'm fully accustomed to clients specifying very exact UI/layout requirements, and ever since Anvil, I've been able to satisfy extremely demanding requests for precisely laid out UIs. And now, with GPT writing UI in Flask Jinja templates, the limitations have been utterly smashed, and there's simply no challenge to it any more.
And then this project achieved the highest level of success I could have imagined.
I've been working on one particularly large, complex project for more than a year, and this client decided several months ago that she wanted to hire her graphic designer to skin the software we'd built, so that it perfectly matched her established brand styling.
Her graphic designer took a number of weeks to build layouts in Adobe XD, which allows artists to build UI layouts visually, and then XD provides all the details of every layout element in the design, as color values, sizes, rounded corner values, downloadable icons/images, CSS styling, etc.
Basically, as a developer, you can hover over anything in an XD design, and access the values, CSS, etc., needed to reproduce the exact layout, so that those values can be manually entered into whatever UI code/framework you're building with - but there's no way to bulk import them into your custom code. There are literally thousands of potential values to be copied into your layout tooling, which could easily take many hundreds of hours of labor to reproduce in a project of this size, which consists of many dozens of screens, each with multiple alternative responsive desktop and mobile layouts for different sized devices, etc.
To be clear, this application consists of several connected applications, each comprised of tens of thousands of lines of code to implement all the functionality (separate from UI styling). And all that functionality needed to remain 100% untouched while applying the new styling work.
I was able to use GPT to first create generic CSS styling for every Jinja template in the application, totally separate from functionality - that took literally a minute. Then I fed screen shots of each XD design into GPT, and it updated the generic styling to perfectly match the specific layout designs in XD - without ever having to touch any code whatsoever.
The great thing is that if my client wants to edit any styling to exactly, perfectly match her graphic designer's work, with pixel perfect precision, GPT is able to do that too, by utilizing any individual values that XD provides (I did quite a bit of this precision work afterward, just to tweak layouts exactly, and to make sure there were no limitations to the possibility of achieving specific detailed goals).
Of course, I can also simply ask GPT in plain English to make any general changes to the look/feel/layout of anything on screen, as I always do. It's just as easy to morph the exactly specified layout values generated by screen shots and values copied from XD (or choose *any other UI design tool), with plain English, and without unintended regressions. GPT 5.2 has become absolutely fantastic, BTW, at avoiding unintended regressions, and only touching exactly what your specify, not just for UI but for all the functionality in an app.
This entire UI skinning project took only 1 long day to complete, from beginning to end, and the next time I need to do this sort of thing it will go *much faster. My client is ecstatic with the result, and her graphic designer is fully satisfied and so relieved that she didn't need to do any coding whatsoever (as am I!).
I thought it would be several more years down the road before we'd get to this level of LLM driven software development capability. And of course this new UI capability is great, but the functionality which can be built - exactly to specified requirements - is far deeper and more complex (consider that the UI of this project took one day to build, and the functionality took hundreds of times longer ).
This is not 'vibe coding' - it's real, careful and meticulously specified engineering, but with LLMs doing all the work. In my recent experience, for the sort of work I do, there are no limitations on what can be achieved, and the tool set & workflows I currently use *far outperform the productivity and capability of anything I've ever experienced in the past.