I'm interested in understanding your desire to do design and prototyping as a single shot?
My expectation was that I'd iterate on a few UX designs with the LLM and then when I'm happy with what the LLM is suggesting, I'd output to figma, and then maybe move to code.
It's great that you're generating code, but isn't that increasing your cost and processing time to write code for each iteration?
What do you mean by "single shot?" Are you referring to one-shot prompting? I should have clarified in my post — we do have customers who one-shot designs - but that's very rare (it's usually landing pages because Sonnet 3.7 is really good at those). We heavily encourage iteration and expect it. The longest thread on the platform is 850+ messages in a single chat!
We have a fine-tuned fast apply model for applying code diffs, so that helps minimize the processing time. Always trying to make it faster though.
We view code as a way to 1) unlock interactivity, 2) communicate with the LLM.
A question for you: if we weren't asking the LLM to generate code, what would we ask it to generate?
1) as for a certain UX design
2) AI shows me an image of what it thinks I want
3) I make suggestions, changes to what I want
4) AI makes changes, shows me an image
5) back to step 3 and repeat until I'm ready to view code
6) have the AI write the code of the UX once.
I understand this may mean there are multiple images showing a flow, or different states, but in my mind, the time consuming part is the AI creating each JS file for the UX. I would think it could iterate quicker on designs and then output code.
This is how UX is designed today, designer does a bunch of iterations, gets feedback, then hands it to developer.
Right now, you're doing this all at once (or at least that is what my experience was).
I get what you mean about code as a method of communicating with the LLM, and I'm an engineer, so I kinda get that, but my first reaction wasn't "let's look at the code output", I was looking at the design output and thinking "is this what I want?"
I don't think generating images would be faster than generating code, plus you lose all the interactivity. Many designers tell us where they get the most value out of Magic Patterns is that the designs are interactive by default, unlike traditional vector-based design tools.
I'm with you that the first reaction is not "let's look at code output." In fact, for most of our users, they never look at the code. It's abstracted away from them and just an implementation detail to unlock interactivity.
My expectation was that I'd iterate on a few UX designs with the LLM and then when I'm happy with what the LLM is suggesting, I'd output to figma, and then maybe move to code.
It's great that you're generating code, but isn't that increasing your cost and processing time to write code for each iteration?