“Crafting the last mile of delight”

What will it mean when an algorithm can handle 90 percent of your job? What will working with dozens of agents be like?  When anyone can build enterprise tools—instantaneously? 

That future has already arrived at the major AI labs—and faster than almost everyone expected, even Anthropic. Founded in 2021 by a group of researchers who set out to build AI that is safe and useful, it’s now used by millions and was valued at $965 billion in its most recent funding round.

The company operates on the assumption that AI capabilities will keep improving quickly—and that its own internal ways of working have to keep pace. That means fomenting and accepting exponential change. Speaking to Joel Lewenstein, Anthropic’s refreshingly honest head of product design, provides a visceral sense of what the future we’re all heading toward feels like—both exhilarating and, occasionally, overwhelming. 

I spoke with Lewenstein about why protecting our thinking matters more than ever, why his team members don’t feel AI is atrophying their craft skills, and why enterprise collaboration is the next frontier for research and product.

Ed White (EW): Do you remember a moment when you first realized AI’s true potential?

Joel Lewenstein (JL):  My wife is an appellate lawyer. It was around 2023, and she was trying to explain this extremely dense legal case to me. After three tries, I still didn’t understand it. So one night, while she was asleep, I pulled out my phone and asked a gen AI to explain the case. It got two-thirds right, but then it started hallucinating, providing incorrect but insistently confident information. I remember having this conspiracy theorist-like out-of-body experience. I felt so alive because I’m learning this deep thing, and I felt so scared and confused because I’m being lied to, but it was insistent. I remember thinking that there’s so much good here, along with confusion and complexity that would be interesting to work on. 

The second moment was hearing Dario [Amodei, CEO of Anthropic] on the Dwarkesh Podcast. This was also in 2023. Dwarkesh asked him: “If we reach AGI [artificial general intelligence] and it cures cancer, should it be governed by a company, the US government, or the international government?” I thought it was a crazy premise. But Dario launched into a deep reflection on governance, ethics, and international order. I realized these people fully believed in the goodness of AI. That made me very excited.  

EW: What’s your role at Anthropic?

JL: I lead the product design, user research, and content design teams. There are about 30 folks, and we basically put a product designer, a user researcher, and a content designer on our platform API business (Claude Code, Cowork), our consumer apps, and our enterprise and growth areas. Our designers are deeply embedded and trying to figure out what the hell this role is in 2026. 

EW: Having come from more classic product design roles at organizations like Airtable and Quora, how is this experience different?

JL: The most interesting dimension here is speed. You hear this constantly, right? But the weird part is the time compression. There’s a measure-twice-cut-once quality to the work. You do a lot more preparation, thinking, and consideration before you build because the time to build is so short. The question is: How do we keep up in an environment where everyone’s shipping all the time?  

We also build a ton of internal tools. Our content design team created a GitHub agent that monitors strings sent to production. It checks for adherence to our content guidelines and opens PRs [pull requests] for any that don’t comply. That’s crazy because this used to be 50 percent of the job. Now it’s just 5 percent, allowing us to focus on other things. All of our designers are writing code.  

EW: Where does the design or craft bit fit in, then?

JL: With design, there are three steps to the process. Step one: determining what we should build, why we’re building it, and what problem we’re solving. Step two: creating user flows and ensuring the basic components are in place. Step three: crafting the last mile of delight, those little details that many people may not notice, but subtly make a difference. Step two is now gone. We just prompt Claude, and it just does it. Step one—deciding whether we should build something, why we should do it, and where it leads—is still a very human and messy process. As for the last step of getting all of those little interactions and visual details right, our designers are still better than the models. Maybe that won’t be the case forever, but for now, it is.  

EW: Tim Brown, IDEO’s former CEO, co-authored an article about this recently titled “The AI dividend.” Now that AI has reduced a task that used to take 50 percent of your time to 5 percent, what do you feel it’s allowing you to do?  

JL: I would love to tell you that all the boring tasks have been automated, allowing the designers to sit at whiteboards thinking bigger, more ambitious thoughts. But our Slack and GitHub are flooded with PRs from dozens of engineers. We’re juggling 17 different Claude Code instances, and each one is pinging, “I’m done! Can you review this?” It’s a cacophony of tiny projects. It’s similar to the attention-economy notification overload we experienced in our personal and social media lives. Now, it feels the same way in our work life. We’re fighting harder than ever for half a day or a day to just think, dream, and explore.  

EW: Do you think the pendulum will eventually swing back? 

JL: It has to because the human condition will require it. I think two things will happen. First, we haven’t invented an interface for managing hundreds or thousands of Claude instances.  There’s still an unsolved UX problem. The best analogy is the American presidency. There are hundreds, thousands, or even millions of staffers running around making big and small decisions, producing piles of status updates. A human chief of staff manages all that. Eventually, we’ll need a digital equivalent who will serve up a concise summary from a few of your SVPs [senior vice presidents] on Monday morning—agents who have dealt with 10,000 individual contributors—so that you don’t have to think about each task.

EW: How else are you seeing designers’ roles change?  

JL: Being close to the code and adopting a code-first, prototype-first mindset is really central. We’re not too precious about our work. If something has a decent UX, we put it out there in a beta and collect feedback. 

As someone who has always worked in the digital space, I think about how creative professionals in the past, like musicians, operated: once an album was finished and the record was pressed, that was it. I could never have worked that way creatively. Back when I started, in the Ruby on Rails era, we could simply deploy again next week and fix any issues. But for AI native folks, if something goes wrong on a Tuesday night, they just submit a PR on Wednesday morning to fix it. Everything is malleable and adaptable all the time.

This lack of preciousness, the idea that “done and shipped” is better than “perfect,” is crucial. Of course, there are challenges. Sometimes the coded output doesn’t look great, but we accept we’re going to learn from this vibe-coded thing and get on with it. We refer to this approach as “intentional craft,” meaning that if we need to make this outstanding, we still have the chops to do it, but we also need the discipline and the judgment to recognize when it’s appropriate to let a rough version go out.

User expectations are changing fast. During the heart of the mobile app era, people had incredibly high bars for their experience. We’re no longer in that era. When we put out these vibe-coded, early experimental things, we get feedback like, “Guys, this broke three times, and this button is called three different things.” But they also say, “I love it. Please make it better, and I'll keep using it.” We’re in a part of the S-curve where users have a lot of tolerance for jank. Eventually, this field will mature.  

EW: How do you maintain your craft in a world where models can perform many design tasks?

JL: Collaboration with these models is not a replacement story. You get better as a writer, editor, and thinker with a creative partner who pushes you, asks questions, and explores alternative theories. There’s a nice sparring partner quality to collaborating with AI that I appreciate.  

The models aren’t doing anything better than a thoughtful designer can with enough time. In a head-to-head competition between Claude and an outstanding designer, the designer still wins most of the time. I don’t hear our designers being concerned about skill atrophy. They’re worried about not having the time and mental space to practice the skills they already have and care about.   

EW: That’s very heartening to hear, to be honest. With so many options of what to build and many more people who can do many more things, how do you find a sense of direction? Do you and Mike [Krieger, co-lead of Anthropic Labs] share an ethos on product development?

JL: We’re letting the model lead the way. For example, in the pre-AI era, a product like Claude Design would have emerged through a TAM [total addressable market] analysis: evaluating our position as market leaders in coding, identifying adjacent spaces, interviewing stakeholders, spinning up a team, and committing to a 6- to 18-month timeline. 

In reality, the complete opposite unfolded. People using Claude Code were organically making UIs [user interfaces]. One of our designers noticed that these URIs [uniform resource identifiers] were pretty good and wrote a skill to coax the models into producing even better front-end designs, specifically using our design system. He realized it was useful and spoke with our researchers, who confirmed it wasn’t a fluke. He then refined the skill into a standalone app. That’s the genesis of Claude Design. We felt comfortable launching it at a research preview because we believe the models will continue to advance. 

EW: That’s interesting because organizations don’t generally use what I’m going to call “academic research” as a way of thinking about how products come about. 

JL: It’s crazy because at any other place that I worked, nothing changes on a product unless you assign an engineering team to change it. If your product engineers all went on vacation, no features would launch. But at Anthropic, if our product team went on vacation, the models would just keep improving.

EW: What’s it like working alongside agents daily, and how does it impact the organization?

JL: Everyone at Anthropic has dozens of Claude Code agents running in the background. We recently worked on a navigation IA [information architecture] refresh. Anyone who’s done a navigation redesign knows there are a million questions in the kickoff meeting: Are people really using this tab? Are the users of this tab and this tab the same, etc. In my old-school, pre-AI brain, I’m like, “We’ll need a week to explore, a follow-up meeting, and then we’ll come back with all these questions.” But in the meeting, the designers are just tap, tap, tapping on their Claude Code agents, asking, “Will you look at our internal documents, Slack, and our data to figure out how many people click on this tab?” And 45 seconds later, someone chimes in, “Hey, it turns out that tab is still used a lot.” It’s like everybody has a team of research assistants behind them, finding answers for them in real time. 

EW: What’s inspiring you right now? 

JL: I’m curious what the next three to five years hold for the unit of computing. We learned how to use websites, then apps, and now chatbots. What will be the next unit, and will it be vibe-coded or developed by individuals, corporations, or models that generate it in real time?

EW: Looking ahead, what’s most exciting to you and your team?

JL: We’re thinking a lot about teams and team productivity. Claude is amazing in single-player mode. But while we’ve become more productive individually, the collaborative aspect of work hasn’t changed dramatically. Right now, we have a World Cup collection of individual superstars producing amazing things, but their collective chemistry remains unmediated by AI. But then again, you and I are speaking right now completely unmediated by AI, and that’s also beautiful.  

EW: What advice would you give designers? What should they be doing and thinking about?  

JL: When users have a problem, our instinct as designers is to create a new feature—design the rectangles and the flows—and let AI build it. We’ve started to invert that approach. If users have a problem, we ask, “Can Claude solve it on its own with prompting?” Then, if it can’t or needs additional structure, we add a feature or product architecture. 

‍

“Design in the Age of AI” is a series of conversations with designers and makers from across industries and disciplines, building the future with AI today.

‍