The only design that can’t really be faked nowadays is large printed uppercase with misspellings and nearly incoherent grammar on a white paper taped carefully to a cubicle divider.
Loving the collage of screenshots where every one screams of being AI generated
The only ones in the readme that don't feel like slop, to my eye, are Cold Snap and Mend Assembly - I like those. The rest are all practically identical and are the go-to output for what you get if you ask Sonnet to build a website.
I have used skills like impeccable, ui-ux-max, and many, many others and they all seem to converge on the same thing. They are good to help out with some inconsistencies, accessibility, streamlining, but the design itself is the same unless you are extremely explicit.
It at least tries to attack the convergence directly by keeping a log of past builds in .hallmark/log.json and hard-fails a run that repeats the last few page structures or themes. While it still admits its nav-rotation rule is "the single most-violated rule in practice"...
I have tried all possible ways: skills, shuffle.dev ai editor, moonchild.ai, giving screenshots of existing sites as a sample. I don’t count Lovable or Figma html2figma plugin or figma2code plugin etc.
After burning a lot of time and endless tinkering, I have realized that the only way to speed up the process is draw on paper into the most detail you can, then redraw into figma or penpot (architects that draw more nicely, can scan images) and then to prompt a LLM (multimodal one that can ingest images).
This may lead to less sloppy design.
I have tried impeccable, jakubkrehel skills, and many others.
LLMs still put something AI-ish into design anyway, you can feel it. So first faster step with less iterations is simple drawings and design logic by hand. (There was a study that said that drawing by hand uses different brain ways than drawing with mouse - this is the reason behind artists’ use of large scale tablets and multitouch pens.)