Claude Design is just a big opinionated prompt: https://www.lobsterpack.com/blog/claude-design-trenchcoat/. Among other things, it knows it isn't that great at drawing SVGs, so it won't try unless you force it to. For a logo, try drawing it with vanilla Claude Code as if it was a separate project: ask a "design agency" to ask you questions, answer them, then make a detailed brief of what you want to draw, then make it output an exact plan of what and where will it draw, lastly ask the chatbot to do the actual drawing using sub-agents for drawing individual components. Also add a "render it to a raster and make sure it looks right" step as well.

> Claude Design is just a big opinionated prompt: https://www.lobsterpack.com/blog/claude-design-trenchcoat/.

Do you think anyone expected it to be something other than a prompt with skills, tools, etc?

Also, you certainly got some brass, linking an entirely AI generated article in a forum where extreme distaste is registered for entirely AI generated posts.

How is your comment not downvoted to oblivion?

Sure, the blog post is AI generated (I'm not a native English speaker, our bunch is often shy about our English language skills), yet the research I've done there is manual.

I found it interesting that (contrary to the popular opinion) there wasn't some magic, e.g. a novel model, happening with Claude Design, especially magic enough that Figma wouldn't be able to replicate if they felt like it.

Also, apparently human (not artificial) neurons were behind that huge prompt, very well aware of the limitations of the model, cheating here and there to make Design's outputs more impressive, making it "create a (design) plan" beforehand, i.e. all the stuff that we the common laymen could do ourselves with the same tools.

> Also, you certainly got some brass, linking an entirely AI generated article in a forum where extreme distaste is registered for entirely AI generated posts.

> How is your comment not downvoted to oblivion?

I'm sure there's a polite way to say things.

I heavily dislike LLM content, but if you read the content, it's actually got information of value.

> I'm sure there's a polite way to say things.

That was the polite way of saying things. The phrase "if you couldn't be bothered to write it why would anyone bother to read it" was a saying from usenet times.

The truth is it took the author less time to "write" that piece than you to read it. It's a blog. There's no deadline, and yet they couldn't take the time to actually type out their own thoughts.

> I heavily dislike LLM content, but if you read the content, it's actually got information of value.

If it was so valuable the author would have written it themselves.

My prediction for every creator tool will be 30% deterministic software and the rest 70 is a long system prompt. This is the future.

Claude Design is very explicitly oriented towards product UI design. It's not trying to be a product that can make you a logo.

You should try the Arrow SVG model by Quiver, should be much better at that sort of thing since it's made for that.

Anthropic has no image generation models, right?

We've got an LLM using CSS and emojis and maybe pelicans riding bikes (SVGs).

Yes, they only do SVGs.

I'm actually glad they're focusing on code, and code adjacent tooling only.