I’ve been using Claude Design for my front ends. The output looks and feels good enough, but the designs often look very similar and generally adhere to contemporary web tropes.
Keen to hear if anyone has had unconventional creative adventures with it.
[Take a look at my portfolio site](https://reebz.com), please view on desktop. This is about 3 weeks of effort to date. It is unfinished, but you get the idea.
Just like SaaS boilerplate from the decade prior, there is LLM boilerplate (since it’s trained on the internet).
So if you put in enough elbow-grease anything is (still) possible!
Even for the large products, figma is not the starting point for new concepts anymore. I start with a quick prototype on dev environment and then share it with designer for further improvements in figma or in the app itself. With every new model or agent improvement, going back to figma for polishing the ui is decllinig.
If we can find a way to keep the frontend code static templates without complex logic, need for this polishing in figma will go away completey as llms can understand it one context window. With the modern frameworks designed for client side rendering, keeping everything in one context is still tricky.
Figma make and gpt designer have a bunch of catching up to do. I couldn’t even import our brand guidelines into make which is already a .fig like what are we even doing here, guys? CD crunches through ungodly amount of tokens and is really slow on iteration but at least you can get some really nice prototypes extremely quickly there. GPT beats any Anthropic models on illustrations so they really should get a grip on multimodal. Overall, it seems like we’re still super early but you can already see glimpses of what may come
Sounds like desk strat RAD work is moving to LLM gen code at JS. My recentish experience of that kind of work has been Athena at JPMC and Quartz at BoA; both Python with functional style via DAG or pixie with py ui framework to match. Which enables quick dev of the parts of trading workflow that don't need to be quick, like booking tools or EOD risk. I know first hand Athena and Qz are crufty when you get into the weeds. The bonsai framework with Elm inspired ocaml impl sounds v cool. So I can see how this approach can accelerate a lot of trading tech dev. But does it have any traction over the hard problems where we turn to C++ or Rust: near real pricing and risk across multiple instruments and markets?
I use the same approach a lot. Before AI I also did this manually. First sit down with a user and just paper and pencil, then hack together a frontend POC / demo, have them play with it and adjust until it works as they wanted.
For me building a quick (not production quality) frontend demo in code was already often faster than getting the right interaction working in Figma. And it allowed to make it fully interactive so you can catch much more edge cases on the UX side.
Now with Claude Code it's even faster to build the throw away prototype. But not a huge difference since discussing with the users and thinking about how it should work is 80% of the time. Claude maybe halves the other 20% compared to quickly doing it yourself. Faster to first version, slower to iterate if it didn't fully get it.
We are doing this on my team (I am the frontend engineer) and honestly I really miss the old way of doing things.
Written specifications are being reduced in favor of these working prototypes, and now there’s this extra cognitive burden of reading the code and trying to determine what were the intended changes, and what’s the slop that needs to be tossed aside.
We also have to figure out, should we take over this generated PR and make any needed changes? Or do we start over from scratch? There’s often a sense of friction either way.
There have been times where a bunch of unintended changes were generated and I took time to port them over on my reimplementation, and then later on it’s “oops! Sorry! We didn’t mean to change that.”
I get it’s empowering but it does take away from some of the joy I used to find in my work and replaced it with some headaches.
Honestly, I'm not really pro or anti llm and I think there are a ton of limitations for using it to generate code, but UI has been probably the only thing I've been able to vibe code. It helps that I've done a lot of UI work over the years, but I think the combination of defects being easily visible through normal usage, the UI being a non-critical component of a system (bugs don't cause vulns or data corruption (usually), combined with the amount of churn that UI's see, make it a somewhat uniquely good candidate for vibe coding. Also a lot of UI toolkits are declarative, and I think language models do much better with declarative code.
In a way it's not much different from copy-pasting components from templates or whatever, just with more customisability. And for stuff that isn't HTML-based like React it does worse. It's also not great at building component libraries, I still write those myself with little LLM involvement, but that makes sense because the architecture is actually relevant with that, unlike generating CSS and xml-derived components, which is mostly just declarative templating anyways.
I've had decent success writing the core logic myself and then delegating the UI to AI. I think if I didn't write the core logic it would not work very well, but since it's designed well by myself the AI has a much smaller scope to work in which constrains it enough where vibe coding works. Pretty cool.
The benefit here is designers learning to code. It was always weird to me that designers were shaping software without knowing how it was built. I'm a designer btw.
However, designing in code is technology-first. One could argue that the purpose of design - to shape the artifacts for human purpose - is better done NOT starting with the strict rules of code. Pen and paper is still hard to beat, not for anything that looks nice, but for helping your mind forward.
I worry about Figma stock, I know some who bought during the IPO who are now underwater. Figma launched their own design agent but not sure how well that's doing.
Even if this is ad only - are we really ready for a rug pull from Anthropic?
Maybe I’m completely unaware in this tools space, but I feel like it’s last tool that’s worth (and not pricey)…
It's much harder to RL out design taste because it's not self-grounding, and human labelers have no real skin in the game, so this (having a human with a vested outcome in the process directing a model's work) is the best way to get LLMs better at design/"taste"/aesthetic judgment themselves. We were working on the same thing 7 months ago and then I realized that winning over designers to do this would be a huge uphill battle setting up an inevitable fall from grace later on.
What makes me most suspicious of Claude Design is that when you disconnect and reconnect later, it loses context and nags you that the product doesn't work like that. Bullshit. It's at best an anti-abuse/implementation detail (to keep you from launching 10 at once and coming back to them later) or product shortcoming that just so happens to be optimized for keeping you from continuing your design in better tools than theirs for the inevitable followups.
It's great for one shots and it makes sense when you're trying to build a vertical product development stack like Anthropic but I'm disappointed it feels more like a tool optimized for keeping you in their product than for what you're working on. If a company other than Anthropic had shipped this - it's not that hard to build a visual self-eval loop, just use Chrome Devtools Protocol to run headless chrome and take screenshots -> feed into a judge LLM for feedback -> continue - I don't think it would really have seen much adoption.
That said, AI trained on Actor-Critic with a tight human feedback loop definitely seems like the right approach to solving the problem, just not something I want to spend my time training for someone else unless I can do so with higher "entropy" ie high parallelism/optionality
58 comments
[ 2.0 ms ] story [ 50.6 ms ] threadKeen to hear if anyone has had unconventional creative adventures with it.
Just like SaaS boilerplate from the decade prior, there is LLM boilerplate (since it’s trained on the internet).
So if you put in enough elbow-grease anything is (still) possible!
from 6 sessions and 5 projects only one template that I choose anything else is really really bad
For me building a quick (not production quality) frontend demo in code was already often faster than getting the right interaction working in Figma. And it allowed to make it fully interactive so you can catch much more edge cases on the UX side.
Now with Claude Code it's even faster to build the throw away prototype. But not a huge difference since discussing with the users and thinking about how it should work is 80% of the time. Claude maybe halves the other 20% compared to quickly doing it yourself. Faster to first version, slower to iterate if it didn't fully get it.
Written specifications are being reduced in favor of these working prototypes, and now there’s this extra cognitive burden of reading the code and trying to determine what were the intended changes, and what’s the slop that needs to be tossed aside.
We also have to figure out, should we take over this generated PR and make any needed changes? Or do we start over from scratch? There’s often a sense of friction either way.
There have been times where a bunch of unintended changes were generated and I took time to port them over on my reimplementation, and then later on it’s “oops! Sorry! We didn’t mean to change that.”
I get it’s empowering but it does take away from some of the joy I used to find in my work and replaced it with some headaches.
Do you not pay for Claude?
In a way it's not much different from copy-pasting components from templates or whatever, just with more customisability. And for stuff that isn't HTML-based like React it does worse. It's also not great at building component libraries, I still write those myself with little LLM involvement, but that makes sense because the architecture is actually relevant with that, unlike generating CSS and xml-derived components, which is mostly just declarative templating anyways.
I've had decent success writing the core logic myself and then delegating the UI to AI. I think if I didn't write the core logic it would not work very well, but since it's designed well by myself the AI has a much smaller scope to work in which constrains it enough where vibe coding works. Pretty cool.
However, designing in code is technology-first. One could argue that the purpose of design - to shape the artifacts for human purpose - is better done NOT starting with the strict rules of code. Pen and paper is still hard to beat, not for anything that looks nice, but for helping your mind forward.
It's much harder to RL out design taste because it's not self-grounding, and human labelers have no real skin in the game, so this (having a human with a vested outcome in the process directing a model's work) is the best way to get LLMs better at design/"taste"/aesthetic judgment themselves. We were working on the same thing 7 months ago and then I realized that winning over designers to do this would be a huge uphill battle setting up an inevitable fall from grace later on.
What makes me most suspicious of Claude Design is that when you disconnect and reconnect later, it loses context and nags you that the product doesn't work like that. Bullshit. It's at best an anti-abuse/implementation detail (to keep you from launching 10 at once and coming back to them later) or product shortcoming that just so happens to be optimized for keeping you from continuing your design in better tools than theirs for the inevitable followups.
It's great for one shots and it makes sense when you're trying to build a vertical product development stack like Anthropic but I'm disappointed it feels more like a tool optimized for keeping you in their product than for what you're working on. If a company other than Anthropic had shipped this - it's not that hard to build a visual self-eval loop, just use Chrome Devtools Protocol to run headless chrome and take screenshots -> feed into a judge LLM for feedback -> continue - I don't think it would really have seen much adoption.
That said, AI trained on Actor-Critic with a tight human feedback loop definitely seems like the right approach to solving the problem, just not something I want to spend my time training for someone else unless I can do so with higher "entropy" ie high parallelism/optionality
Klankers will fix everything. Right?