Sam redesigned an internal web app in Claude Design and handed it to Claude Code, live in 30 minutes. Plus how projects start, ad concepts and Figma.
Claude Design is Anthropic's chat-based design tool, and in Sam's hands it went from a prompt to a fully redesigned internal web app running in production in about thirty minutes. On Episode 40 of Built This Week, Sam Nadler walked Jordan Metzner through three projects he built in the first days after launch: Meta ad concepts, a shared presentation for the marketing team, and an app redesign he handed straight to Claude Code. We also covered SpaceX's option to buy Cursor, ChatGPT Images 2.0 and why AI labs go quiet between model launches.
Sam says he has seen mixed reviews of Claude Design on social media, but his own experience was great: high design quality and easy to use. To him it feels like having a UX and UI designer at your hip that you can chat with, engage with and ask to propose ideas. The output is high quality, and it gives you a lot of options: you can change the font and color, or keep chatting.
Every project starts with a prompt, such as building a deck, redesigning a web app or making social ads. The first thing Claude Design usually creates is a questionnaire of eight to 12 questions. Once those are answered, it produces the first set of assets with quick tweaks you can make in the moment, and everything else is managed through the left-hand sidebar. Sam also noticed it prompting him to start a new chat to save on context, so he opens a new chat whenever he moves to a new part of a project.
Watch the Claude Design handoff to Claude Code in Episode 40 (starts at 5:44) · Episode 40 notes and transcript
Yes. Sam asked Claude Design for HipTrain Meta ad concepts and then told it to target women between 50 and 70, one of the service's prime demographics. He wanted the tone a little cheeky and self-aware, and he says it captured that to some degree. It built a presentation with three campaign concepts and laid out how it thought about each one.
He also asked it to write an image prompt for each campaign, plus a Veo 3 prompt in case he wanted to turn one into a video. For one concept he took the image prompt, ran it through Nano Banana, and gave the image back to Claude Design, which placed it into the finished ad. From there, he says, you can keep iterating.
Jordan asked how you actually get the work out. Sam says there are many share options. The one he found most useful was a link that turns the project into a presentation his marketing team could comment on, or even edit directly inside Claude Design. You can also download the project as a zip, export a PDF, send it to Canva or export standalone HTML.
The option that stood out was "hand off to Claude Code." For a website or web app redesign, you can link your local code base to Claude Design, work through the design in chat, and then click the handoff button. It creates a zip file that you give to Claude Code. Sam did this with an internal web app his team uses. Claude Design gave him a polished redesign page by page, keeping the exact functionality of the app.
I then gave it to Claude Code. And within, however long it took, it actually took quite a while, thirty minutes, the entire web app was fully updated to the new design and the results were amazing.
— Sam Nadler, Episode 40
He has since shown teammates who maintain their own web apps how to do the same thing, and an hour before we recorded they showed him their new designs. The full walkthrough is on our internal web app redesign build page.
Jordan asked whether Sam would abandon Claude Design if these features were simply rolled into Claude Code. Sam said yes, maybe through Cowork, and that he was a little surprised it shipped as a separate product. Jordan's read is that it was a deliberate test:
it's just much easier to build a web app than it is to integrate something into a native desktop application.
— Jordan Metzner, Episode 40
In his view, a web team can move fast and check whether customers like a product before it gets rolled into the desktop app, as he thinks happened with Cowork, which we built with live on Episode 29. He called Claude Design "a little bit of a Figma killer," while saying he wasn't sure. Sam agreed it is far easier:
I would say my Figma skills are pretty awful, and this is incredibly easy.
— Sam Nadler, Episode 40
Sam thinks Figma will keep its place, but for him Claude Design felt like a step function in what he can do. We compare the two in more depth on our Claude Design vs Figma page.
Sam said he didn't know the exact deal terms, but his understanding was that SpaceX has the option to buy Cursor by the end of the year for $60 billion, or pay a $10 billion breakup fee. SpaceX has the compute and Cursor has the coding agent.
Jordan set the context: most people saw Anthropic leading the coding race, OpenAI second, and Google and Grok behind. Elon had said Grok was behind in enterprise coding, and had recently set up one of the largest GPU data centers with NVIDIA. Reading the deal, Jordan sees xAI and SpaceX, which have now been merged, offering excess compute to Cursor along with an option to buy. That gives SpaceX room to learn whether Cursor's next step is about inference or about a new model to compete with Claude Code and Codex.
He noted public data put Cursor at about a $2 billion run rate, so $60 billion is roughly 30 times revenue. The most advanced AI teams have moved on from Cursor to Claude Code and Codex, he says, but many enterprises and universities are only now adopting it. His take is that the deal may put SpaceX, Grok and Cursor back into the number three slot against Google, and make token prices more competitive. Sam's response: it still feels like the first inning of a very long game.
ChatGPT launched Images 2.0 the day before we recorded, so we had both had limited time with it. Sam says the reports, mostly on social media, point to much better text fidelity. Older models would turn a word like "Jordan" into gibberish, and he found the new one pretty good.
Jordan was underimpressed. The background behind him in this episode was made with it, but he couldn't see a significant difference from the backgrounds he used to make with Nano Banana, and he felt the launch examples were better than his results. He still likes the competition: the more models there are and the cheaper images and video get, the better.
Sam felt Google had gone quiet over the last quarter while Anthropic took the lead. Jordan answered with an analogy: training a model is like building a hotel. A hotel takes years to open, and by then demand has grown and you need another one, so you are always a bit behind.
if it takes three months, six months, nine months to train some of these models, then it seems obvious that you're gonna go through these quiet periods where it looks like some of these major laboratories are doing absolutely nothing.
— Jordan Metzner, Episode 40
Behind the scenes, he says, labs are training at maximum, and the hard call is when to start the next model, since a training run is not easy to stop or change once it begins. As models get heavier and take longer, he expects more of these lulls. He also pointed out that Google had just announced two new TPUs, that Elon tweeted TPUs are underrated, and that Google is becoming a major inference player with new inference deals. Our conversation on inference costs in Episode 37 goes deeper on that side.
Sam shares his screen through the ad concepts, the share menu and the before-and-after app redesign. Watch Episode 40 of Built This Week, where the full transcript is also on the page.
