Since Karpathy coined the term vibe coding, some of my non-coding friends have been showing me their work and teasing me that they can now do everything I can. Anthropic just dropped Claude Design. Now it's my turn.
Shelfcritter, the polygon
I've been building Shelfcritter - a book management app - on and off for a while. It's not really about the books. It's my polygon for testing different agents and workflows, letting AI drive parts of the codebase without losing the plot. And it works, mostly. You can add a book, track what you've read, the basics. But every time I opened the thing, I'd wince. The look and feel was off. Here is how it used to look:
So when Claude Design dropped, I knew exactly where it was going first.
Five minutes and 25% of my weekly Design quota
I didn't bother with the tutorial. The UI is intuitive enough that I just started: took a couple of screenshots of the parts that bugged me most, dumped them in, listed the pain points, and asked for two things - a design system and a redesigned flow for adding new books.
Watching it reason was the fun bit. It picked up friction points from the screenshots without my help, then came back with a couple of sharp clarifying questions. I answered. Five minutes later (and about 25% of my weekly Design quota on a Max x5 plan) I had a clickable prototype I actually wanted to use.
Getting the design into the app was the next question. Claude Design has a pile of export options, but for me the obvious move was Claude Code. The handoff is dead simple: Claude Design hands you a prompt with a link, your Claude Code instance fetches it and reads through the design tokens and component hierarchy on its own. No copy-pasting CSS variables. No reverse-engineering spacing from screenshots.
A few prompts later I had a stack of Beads issues, and my parallel-agents workflow did its thing. Shelfcritter ended up with proper typography and a sense of calm to it. A real library theme. Not production-ready yet, but a couple hours in, it was close.
How Shelfcritter got AI Digest
A bit less than two months ago I started running a daily AI digest. The premise was small: keep some friends informed about what's actually moving in the AI world, and test out Cowork for automation while I was at it (I'm too much of a chicken to give OpenClaw all the permissions it wants - in theory it could spend all my money or get in a fight with my boss). The digest was the first idea that came to mind when I went looking for something to put Cowork through its paces, and it stuck.
It all runs as a Cowork scheduled task. The flow: Cowork gathers the data → Obsidian vault → Discord. It checks the day's releases and the AI news, writes the whole thing into the vault with proper tags, and posts it to the channel my friends actually read (or I hope they do). A single prompt in the repo walks Claude through every step. Most days I don't touch any of it. However, it is ugly and one-dimensional. Like an endless pit of information, going to that channel to be lost forever.
The $100 that pushed me into building an archive
Then Anthropic ran a promotion and I picked up a $100 offer. I decided to go a little wild: a public archive site for every digest I've sent and a future ones, so the Discord posts aren't the only place this stuff lives.
I didn't really write the site. The same Claude Code workflow that redesigned Shelfcritter - Beads issues, parallel agents, reviewing one phase at a time - is what actually built it. I scoped it out, fed the agents one phase, reviewed the diffs, course-corrected, next phase. Eight phases in, the site was there.
The stack is small on purpose: Astro on Vercel for the static build, Neon Postgres with pgvector for hybrid search, Jina for embeddings. The Obsidian vault is the source of truth. I write a digest, push to main, and the rest just happens.
Claude Design, warts and all
It's not all roses though. Claude Design is in what Anthropic calls a "research preview". You and I can call it beta, like normal people. And it shows.
Twice, while I was working on the archive site, it ate the search page. Not errored. Not crashed. Just gone. There's no checkpoint system, so it wasn't coming back. I regenerated it from scratch both times. And since regenerating costs money, it ate that too.
A few times it also flat out ignored my instructions. Probably a skill issue on my end, but the odd bit is that the thing it did instead was often good enough that I told Claude Code to ship that instead. Pro or con? Still undecided.
None of this is a dealbreaker for me. This kind of stuff piles up angry users fast though, and I hope Anthropic patches it before it does.
And yes - it burned through $70 of that $100 credit in a blink of an eye.
Are designers cooked?
Resounding no.
I'm an experienced UI developer. That's what let me steer Claude Design to the result I wanted. I could tell when it was drifting and push back. Sometimes I'd even catch it doing something better than what I'd asked for. Plenty of people will get decent results too. But for anything non-trivial, an experienced designer using Claude Design will still have an edge. They'll obliterate everyone else when it comes to web design.
Web. That's the key bit. I also pointed it at another pet project of mine, a Beads explorer in Swift. The results were nowhere near stellar. It ignored Apple's UX conventions completely, and I couldn't get Claude Code to retrofit them either. I ended up with a fine-looking app full of components built from scratch in places I wanted native ones.
Wanna see the digest site? Head to aidigest.shelfcritter.com.
![Defluffer - reduce token usage 📉 by 45% using this one simple trick! [Earthday challenge]](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fiekbgepcutl4jse0sfs0.png)
