1. Podcasts
  2. Behind the Craft
  3. Full Tutorial: How to Fix AI Slop with Claude Design in 40 Minutes
Behind the Craft

15 Nov 2026 · 42m · 18 chapters

Episode artwork

Full Tutorial: How to Fix AI Slop with Claude Design in 40 Minutes | Nate Parrott

In short

How Nate Parrott uses Claude Design (with Claude Code and Claude Artifacts) to rapidly create UI designs, slide decks, dashboards, and motion videos—turning “AI slop” into usable output through heavy context, iterative back-and-forth, and code-based regeneration.

Guest backgrounds

Nate Parrott is a designer at Anthropic behind Claude Design and Claude Motion. He previously worked at The Browser Company (built the Arc browser) and has an engineering background.

Key claims

“You get out what you put in.” Provide screenshots/mood boards, past work, and ask Claude to generate many options (e.g., 10). When designing from an existing product, have Claude read the codebase and recreate the UI on the canvas. Iterate by requesting targeted changes (backseat driving) and using comments on artifacts. Motion quality is improving; current models are “okay,” not great.

Notable examples

Duet ideation—Claude read the codebase and produced multiple dashboard-style variants for a “daily briefing” UI; later iterations made a denser “list on card” design. Dashboard demo using NYC MTA ridership CSV via Google Sheets MCP, with live-updating artifacts. Motion demo using Cloud Motion to animate the dashboard with storyboard + style options, plus timing guidance (“humans can only focus on one place”).

Written by AI. May contain mistakes.

Chapters

0:45

Getting the Best Designs from Claude

Nate shares techniques on how to leverage Claude for effective design.

“Really excited to have Nate show us exactly how he designs real anthropic features, including Duet, I believe, dynamic dashboards and motion videos, and even more.”
5:30

Designing with Claude: A Real Case Study

Nate discusses his experience designing features for Claude, including workflows and iteration.

“So why don't we have you share a screen because I'm super curious how a real anthropic designer actually builds real cloud features.”
14:00

Building Improved Design Infrastructure

Learn about the development of a new infrastructure for design projects.

“that would let us deliver better design quality, then we would do it.”
14:46

User Experience Challenges with Claude Design

Explore the challenges users face with separate design tools and chats.

“And that was a lot of people were saying, hey, why is this a separate thing?”
15:47

User Demographics and Design Tool Accessibility

Discover the diverse user base and their varying familiarity with design.

“Probably just people making slides or videos and stuff.”
16:48

Iterative Design and Feature Integration

Understand the process of iterative design and integrating features into main tools.

“And so the default experience in the standalone surface was we're going to ask you maybe eight questions.”
17:58

Feedback Mechanisms for AI Model Improvement

Learn about providing feedback to improve AI design models and their performance.

“I think the problem comes when people don't actually merge it back into the main service or just start adding more and more stuff to the main service.”
19:10

Training AI to Avoid Common Design Pitfalls

Explore methods to train AI models to avoid biases in design choices.

“I mean, they know about these issues and it's an active area of research to figure out how to avoid them.”
20:56

Advancements in Motion Graphics with AI

Discuss the progress in AI motion graphics and its capabilities over time.

“Or you'll bring your existing brand guidelines.”
22:45

Creating Interactive Dashboards with Claude

Learn how to leverage Claude for building interactive data dashboards.

“So we released a new version of Cloud Artifacts recently.”
Show all 18 chapters
24:29

Expanding Cloud Capabilities for Data Insights

Discover how to expand cloud capabilities for enhanced data insights.

“I could have it make custom visualizations.”
26:21

Crafting Visual Stories with AI Motion Artifacts

Explore the process of creating visual stories using AI motion artifacts.

“is sharing a sensor wheel of the stats, right?”
28:00

Iterating on Animation Design

Learn how to enhance AI-generated animations for better clarity and engagement.

“And from there, I got a bunch of different versions.”
30:20

Creating Compelling Sizzle Reels

Discover the process of creating engaging video content using AI and design principles.

“I guess point, I can tell it what I want.”
32:00

Evaluating AI Design Outputs

Explore effective methods for evaluating AI-generated design and improving outcomes.

“But really it's just anybody's dumb idea can be presented in that format.”
34:00

The Concept of Taste in Design

Discuss the evolving nature of taste in design and the potential impact of AI.

“and a lot of them are like yes-no evals.”
37:00

Working with Cloud Tools for Design

Gain insights on effectively using Cloud tools to streamline design processes.

“And I honestly, whenever I get a typo now in an email, I kind of value it more because it's like proof that there was sort of like proof of work behind it.”
39:20

Reflecting on Product Legacy and Innovation

Reflect on the legacy of the Arc browser and its influence on modern browser innovation.

“from every time we've interacted and every time it's interacted with anybody else in the company.”
A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On Behind the Craft, what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00Oh my god, the user asked me to make an app. I'm gonna invent the most brilliant visual style. In Claude's mind, it thinks this is totally original. You get out what you put in. Give me like 10 different options. Different fonts, different colors, different visual styles. We spent like a year in pre-production before we released ARC. Now at Anthropic, you can kind of do that in one to two days. Cross the threshold from like hilariously bad to just bad. And then from bad to like, this is okay. I don't know the first thing about After Effects. And so it's like, for me, I'm like, oh my God, I have this new capability.

0:32Claude's going to be like, this is going to be like six weeks work. And I'm like, I hate to break it to you, Claude, but we're going to do this together in like the next 20 minutes. So you better get cracking.

0:42Hey, everyone. I'm really excited today to welcome Nate, the designer behind Claude Design, Claude Motion, and more. Really excited to have Nate show us exactly how he designs real anthropic features, including Duet, I believe, dynamic dashboards and motion videos, and even more. Welcome, Nate. Good to be here. Awesome. So maybe let's start by talking overall about Claude has come a long way. Opus 5.5, I love it. It's my favorite model. But how do you get the best designs out of Claude? Like it still kind of adds some of the Claude-isms and stuff. So how do you personally get the best designs?

1:13The thing that I always say is you get out what you put in. And so I always start by giving Claude a ton of context about what I'm designing and how it should feel, how it should look. If I'm designing something totally new from scratch, I'll typically give it screenshots, kind of like a mood board. I'll give examples of other pages that I think are interesting. I'll give it past work of mine that I think is cool. And I'll also ask it to go really wide and give me tons and tons of options. This is how you escape the kind of like Claude, everything looks the same vibe when you're starting from scratch is you say, hey, Claude, here's what we're doing.

1:49give me like 10 different options, different fonts, different colors, different visual styles. And then typically I'll look at those and I'll say, hey, none of these are quite right, but I like a little bit of the first option and I like a little bit of the third option. And so give me like five more options that are different, but based on those. And I'll go from there. If I'm designing from an existing code base and I have an existing product, which is most of what I do. Yeah. I'll give Cloud access to the code base. So now you can use Cloud design from within Cloud code. And so I'll say, hey, Cloud, just read the code base, recreate it from the code by reading the code as a cloud design.

2:24It's pretty good at that now. And then I'll say, okay, from there, let's do some riffs. And because you're working from an existing set of context, you have something that will look like that. Got it. Okay, so maybe even building cloud features, you don't actually need a whole design system or everything. You can just have it read the code and then figure it out. That is how I tend to work these days, which is great because when I was designing with legacy design tools, I would spend a lot of time taking screenshots and recreating them in that tool because that code is a source of truth. And now I can just let Claude do that work.

2:56And Claude is not, for a while, it was pretty bad at it. And now it's pretty good and it's getting better and better. And I think we're a couple of models away from it being where I can just say, hey, look at any code in any language in any environment and recreate it pixel perfectly on the canvas and it'll do it. And when you're starting from scratch on something new, like maybe not an anthropic feature, Do you just be like, hey, just give me like five random designs as different styles as possible? Or you have a random number generator or something? How do you do it? I generally just ask for a ton of stuff.

3:26Sometimes I will actually start by asking for wireframes. So if I think it's a particularly challenging design and I don't want Claude to lose itself in the visual details of what fonts and colors am I going to use, I say, hey, Claude, make a bunch of wireframes with different layouts. And actually, sometimes I'll even ask it to think about existing UIs in the same space and think about how those things are laid out and use those as inspiration. I think what good designers will do is they'll say, hey, what's the problem we're solving here? And what are familiar interfaces that people already understand to give us a shortcut to getting to a level of getting to a product that people already know how to use?

4:03And you can ask Claude to do the same thing. Say, hey, if we're building a to-do list or an inbox, think about other inboxes people use. find one that's the right fit for this product and wireframe something in the same type of structure. Awesome. My personal experience, I had Cloud make some slides for me and the first pass was way too corporate looking. So I just told Cloud, okay, your slides are too corporate looking. Use your creativity as much as possible. Like, you know, you have dynamic HTML, make it look more like a movie than a slide. And I did a really good job after that. Exactly, exactly.

4:33And typically if you're going to do a really big deck, I'd suggest doing that upfront and just saying like, hey, Claude, I want to explore the visuals before you go and build the full deck. And that'll save you a lot of tokens because if Claude makes the entire deck in a style you don't like, then you have to go and tell Claude, hey, redo it in a different style. You're going to be spending all those tokens again. So much better to do it up front. This episode is brought to you by Riverside. I record all my podcasts in Riverside, and they just launched four AI agents to help with video production from start to finish.

5:04You can direct them to prep questions for your guests, tighten the edit, pull out clips, and get your posts ready to publish on social media, all without leaving the platform. Each agent has its own specialty, and the more you use them, the more they remember how you like to get things done. You still get the final say on what goes out. Try all four agents using the link in the description and my code Peter Yang to get one month's free. Now back to our episode. So why don't we have you share a screen because I'm super curious how a real anthropic designer actually builds real cloud features. Yeah, we'll love to see it.

5:39Yeah. So this is a little bit of an anecdote from my time designing on Duet. I was not super involved in the Claude Duet project, our personal agent that we launched the other day, but I helped out a little bit on early ideation. And here's an example of a little thing that I did and just handed off to the team that we didn't end up building, but it was an example of the type of work that I will do and give to folks. So it started with an idea, which was that we want Duet to be proactive and we want it to give you a bunch of ideas for what you can do with it. And we probably have to show you like 20 ideas to find one that is actually something you want to do or have taken off your plate.

6:21And so I was looking for ways to do that. I called the feature, here's what I can help with. And I actually started working on it on the subway coming home from work because I had this idea. I was standing on the platform. I started typing it into my iPhone notes app and I put in a bunch of ideas and thoughts about what I wanted at a very, very high level. So it's a pretty long note and not that long, but there's some stuff in it. I had some examples, stuff like that. And then as I was walking home from the train, I just pasted that into Cloud Code Remote. I put it in our real code base and I said, Hey, Claude, I want you to take this note, come up with a bunch of ideas for what this feature could look like and start from a place of reading our code base and recreating it, like I said earlier.

7:03And next morning, came back to my desktop and got to work. And I had, this is kind of where I was starting with. So Claude had read the whole code base. You can see on the left, it had recreated what we had currently, which was just a daily briefing with a stream of messages. And it had given it a bunch of different options for what it might look like in this new design, right? One where it's highlighting in like one big paragraph and it's highlighting the ideas in line, different visual styles for that. A couple that are just big lists on a card in various styles. And there were more in here that I wasn't able to pull out here.

7:43But that's kind of where we started. And from there, I just went back and forth with Claude and asked for a bunch of like high level changes. So I was like, hey, can we do a version of the list on card that's much denser, no subtitle, allow it to wrap, remove the dividers, please. Kind of like giving high level feedback, like kind of like backseat driving. So typically in cloud design, my work is a mix of backseat driving like this, where I'm just like, change the background colors, reduce the left padding. And sometimes I will go into the design directly and edit it myself because we have pretty full featured controls for going in and editing stuff there.

8:16But because I was doing this somewhat from my phone, and I was just feeling kind of lazy that day, I did a lot of it just through typing, you know, add some examples of recurring tasks, blah, blah, blah, blah, align the icons to the cells, et cetera, et cetera. And wound up with some stuff that was pretty decent, sent it off to the team with screenshots. And if I was the one building this for real, which I was not in this case, but if I was, I would literally just go into the same chat and I'd be like, okay, Claude, time to build it because we're already in the code base. We're already in a Claude code session that is attached to that code base.

8:49And if I tell Claude to go build it right from the designs, it's going to know, okay, I knew exactly what CSS styles, I knew exactly what sizes and fonts and colors and layouts we want in this design. So I'm just going to go and implement those directly in the code base. And there's going to be very little lost in translation, which is really, really nice. And what was the initial conversation that made these five different designs? You had your Apple notes, but did you actually ask for five variations? I believe it was this. There might have been a little more of this, but probably Probably not.

9:17I think it was entirely just pasting this in. And then at the bottom, it was probably something like, just give me a bunch of options for what this looks like. This is typically how I do stuff. I'm like, recreate from code base, give me a bunch of options, and then we move on. Got it. And just to make sure our audience understands this, you have five variations. And then you started iterating on one of them. Did you pick one to start working on? Yes. So I believe that I was actually iterating on a bunch of them in parallel, but I was asking for changes to particular ones at a given time. I'm kind of a hoarder, so I will keep all of my designs on the canvas so I can show people, hey, here's all the different directions that I explored.

9:55But I was typically saying, I think you can see in the screenshot, I'm saying like, you know, can we do a version of the list in card option that's much denser? And so Claude would just copy that on the canvas and make a new version that was denser, but keeps the old version around for comparison. Dude, I love this prompt because even like a year ago, I used to tell people, hey, you got to craft a very detailed prompt and you got to do everything, right? And then just like you, I got more and more lazy. So every single prompt I have, it has many mistakes. And just like, it doesn't actually make sense to humans, but like to Cloud, it actually makes sense.

10:24I mean, recently I've started doing a lot more voice notes, which I think make even less sense to humans because it's just me rambling and yapping. But it's the most efficient way to get ideas out of my head and kind of shape them while I'm working. And so, yeah, it's worked pretty well. And you were able to do this from your phone, basically. So I have used Cloud Design and Cloud Code, but it was on my desktop. So you can actually look at the canvas from your phone and see what's going on? Yeah, you can. You can't do direct editing on your phone, but you can actually comment. So you can pick out an element and point at it and say, fix this, fix that.

10:52You can use interactive prototypes on your phone. It's the same experience. It's just dialed back a little more to fit on your phone screen. But yeah, you could do it all from whatever platform you want. Wow. Okay. So what typically happens, I assume, is you explore some design, you send it out to a team. Maybe someone on a team prototypes it and then you guys all play with it to see how it feels. Is that how it works? Yes, exactly. So with Duet, I don't think we ended up prototyping this exact feature, but the entire way that product was built was people were coming up with ideas, putting them into the internal build of the product, and then just soliciting feedback from people and iterating really rapidly and trying to figure out, okay, we built this feature.

11:28It didn't quite satisfy the need that we wanted to solve, so we're going to try something different. Or we built this feature and we thought it would be cool, but nobody used it, so we're going to try something different. And in my experience, that's the best way to land on good stuff really fast. Oh, when you say nobody uses it, you mean like internally, the ones using it? There's no internal adoption? Yeah, I mean, sometimes you'll build a particular feature as part of a product, and you're going to be really excited about it. And then maybe it's a button or a different way of doing things or a different subscreen or setting or option.

11:57And then you find yourself actually not using it in practice. And so you say, okay, cool, we need a different tack, or maybe this is not so useful. It's kind of crazy because I've been a PM for a long time too. and all the designers I work with, we used to do design sprints and try to do a lot of prep work before you get into the code because code was expensive. But now code is the cheapest thing. It's the cheapest thing. So now it's just like, do you find your job more fun? Because I'm sure you've been a designer for a long time too. Is it more fun now? Just prototyping stuff? I mean, I think it's a lot of fun.

12:25So I'm a little weird. I have an engineering background and so I would actually, I've been doing this the whole time. And before I worked in Anthropic, I worked at a company called The Browser Company where we made a browser called Arc and we worked the exact same way. It just took us a lot longer to build prototypes, but we actually built the same number of them. We spent like a year in pre-production before we released ARK because we had a ton of different ideas and it took us like a couple of weeks to build out each one and see which worked. And now at Anthropic, you can kind of do that in one to two days.

12:53So you've really compressed the cycle of building. What you haven't compressed is the cycle of putting it in people's hands and understanding if they're going to use it, which does take a little bit of time. Yeah, because human attention is still sparse. It makes that more sparse than ever. All right, let me ask you some hard questions. So I think Cloud Design, you built it first as a browser experience, and now it's part of the Cloud Code app, which I think totally makes a ton of sense. Walk me through the rationale behind shipping it as a dedicated experience first, and then merging it back into the main thing.

13:22Yeah, I think it's really easy to iterate in a self-contained space. So when we first built this tool, it was just an internal prototype that I had built mostly for myself. And so I was like, of course, I'm not going to do all the work to make this part of the core product. I just want to have a little tool that will make me designs and slide decks. And if it's really useful for me, it's going to be okay for me to go to a separate surface. And I'll be willing to do that to get value out of it. And that worked really well for quite a long time. It let us iterate really, really rapidly on the experience internally.

13:54And it let us own the entire stack. So we were able to, if we had an idea about how to change the harness or how to change how we call the model, that would let us deliver better design quality, then we would do it. And it would be super, super easy to do and nobody was standing in our way. We rebuilt the entire infrastructure for like how chats work and how you organize projects and how you organize your outputs. And we built the first multiplayer chat within Anthropic, which was kind of wild, just because we were running in the direction of whatever would make our own product as good as possible without having to worry about stepping on the toes of everyone else at the company and hurting the core user experience, which is working with Claude as a general purpose agent and not doing design.

14:36And so that got us a really long way. And as we started burning down the biggest problems that people have with Claude design, one kept sort of, as we eliminated other ones, one of them started moving up. And that was a lot of people were saying, hey, why is this a separate thing? Why do I have to learn a different tool? Why do I have to be in a different place? Why doesn't this work on my phone? Because Claude code works on my phone. Why can't I have my computer use connector in the same chat as a cloud design chat? Why are they separate chats? And so at a certain point, that became our biggest problem.

15:07And we said, okay, this is amazing. We know what we have to do. We've already done all the wide ranging design exploration that was enabled by being in our own really separate surface. Now we can say, okay, we know the product formula and we can invest the time in bringing it into the more polished surface that we all use all day long. And it's a lot more work to build a product, build a feature within a larger product that has way more users because you have to be a good citizen. You can't put your own buttons everywhere. You can't change how the harness works, all that stuff. But we were confident in our ability to do it and that it would be a good thing to do because we'd spent all this time reaching product market fit in the external space.

15:46And I imagine a lot of the users are probably aren't even designers, right? Probably just people making slides or videos and stuff. Yeah, yeah. It's a fair mix. You have a lot of people who are like, I do design, and so it's a faster way to do design. And then you have a lot of other people who are like, well, I actually couldn't use any of the previous generation of design tools, but now I can. So you had a lot of PMs. You had a lot of go-to-market folks. You had a lot of people making slide decks who could make slide decks, but really were not interested in the work of applying the company's design system.

16:15They just wanted to focus on what they had to say. And so you got a lot of folks who are not typically designers. And I think one of my favorite features of cloud design, I've used a lot of AI design tools, but one of my favorite features is after I ask you to do something, it actually shows me a page of a list of questions to clarify things, which is what a good designer would do. But for some reason, the other tools just don't do it. So that's a really brilliant thing that you built. Yeah, we were always very wary of that because I think traditional product design is like, oh, don't add friction.

16:42And what we found actually was that we could not ask too many questions for people. People just loved it. And so the default experience in the standalone surface was we're going to ask you maybe eight questions. And you can skip it if you want, but people actually love to give that input. And it's much easier to do that by being asked than to think upfront about, okay, what questions would a designer ask me? And also a problem that I have, we talk about how just updating the code is much faster than updating the canvas. So for example, I'm trying to build my little fitness app for myself and I just add features left and right to the code directly.

17:15And then do you have any tips for how to keep the canvas updated as the code is updated? I mean, my big one is just when you want to do a design, just ask Cloud to read the code and recreate it from wherever you are at currently. So I don't have files that I keep that are super long running. When I want to design a new thing, I just say, hey, Cloud, read the relevant parts of the code base, create them on the canvas, and then design from there. And very rarely am I coming back to that file and keeping it up to date over a long period of time. Okay, so please just get Cloud to update it, keep it up to date.

17:43Yes. Okay, just to wrap up the Cloud design part of things, I think it's actually a really smart pattern to, like you said, iterate in your own dedicated service first and make changes, and then eventually move it back into the main service. I think I've seen this with co-work and chat too, right? And eventually with Duet, maybe at some point. I think the problem comes when people don't actually merge it back into the main service or just start adding more and more stuff to the main service. Yeah. Exactly. Do you guys have a lot of principles around that stuff? Because I won't name them, but some of these other chat apps just become overly complicated.

18:13It just becomes super confusing experience. Yeah, I mean, I think our general principle is we try not to block new experimental features based on, oh, they have to be integrated. We're okay from time to time launching things that are standalone with the expectation that if they take off, that you'll actually merge them back in. And we treat it as the right thing to do. And so when we said, okay, we're going to go and bring Cloud Design back into the product thing, the product design team was very happy about this because it's like, oh, obviously they should have been the same thing from the beginning, but it makes it so much harder to launch a product from zero to one when you also have to worry about figuring out exactly how it fits into the ecosystem.

18:54Yeah, that makes a lot of sense. One last question on this. So I'm not sure if you still does this or not but Opus used to add these little sidebars to the design and stuff like that. There's like a very claudism stuff. Like how do you give feedback to the researchers to train a model out of this stuff or is that kind of what you do? Yeah, yeah. We give a lot of feedback. I mean, they know about these issues and it's an active area of research to figure out how to avoid them. And it comes from the fact that the model, when you ask Claude to create something, Claude is operating in a fresh context where it thinks really hard and it's like, oh my God, the user asked me to make an app.

19:25I'm going to invent the most brilliant visual style. It's going to be, you know, serif text and it's going to be like an off-white background and it's going to be orange and there's going to be these cool bars on the edges of the cards. Because in Claude's mind, it's only ever done this once, right? And so it thinks this is totally original. And so actually one of the things in prompting that we found to be really valuable is just telling Claude what its failure modes are and what kind of modes it collapses to. So actually in our Claude design prompt, we're like, hey, Claude, for this particular model, you tend to overly bias towards these two fonts and these two colors.

19:59So don't use them unless the user asks you to or you think it's really, really relevant. And that tends to actually help a lot just because the model has no idea what its own biases are. Yeah, I have a no design slot prompt that tries to get rid of a bunch of the stuff too. But like, how do you kind of get the model itself to be aware of the stuff? I guess this is like a longer process, right? You got to like train a certain way or like post train it to avoid stuff. Yeah, I think it's an open question. And I can't talk about what's happening on the research side on this, other than that folks are certainly aware of it and working on it.

20:27But it's an open question. One thing that we've thought about a lot is it has a particular style in which it writes. For a while, it wasn't so good. Now it's good again. But it's a particular style, right? And so when you ask Claude to do a design, should Claude try to invent a design language from scratch? Or should Claude stick to something conservative that works for everything? We don't really know. But one thing I'll say is that hopefully, I think if you're doing design with AI well, you'll kind of make that question irrelevant. Because you will work with Claude to develop an identity that works really well for your particular product or your particular brand.

20:59Or you'll bring your existing brand guidelines. And so you won't have to worry about whatever the model's tendencies are to do when it doesn't have anything to work with. Yeah, that makes a lot of sense. I think your personal context, your brand context is the durable asset that you bring to new models and stuff. Yeah, makes sense. All right, dude. Well, I think recently with Opus 5.5 and Sonnet, like the motion graphics and the video capabilities has blown me away. And did you guys specifically train for this stuff? Or what was the surprise internally too? How amazing? I'll tell you, I've been trying to get call-out to do motion graphics for a long time.

21:30So I remember back at the end of last year, I was trying to make Claude make little videos and little sizzle reels and animated product demos and things like that. And it could barely do them, but it could do them. And they were hilariously bad. Like you look at them and you'd be like, nobody's ever going to look at it. Like this is terrible. But then over the past year, every model that comes out, it's gotten a little bit better and a little bit better. And at a certain point, it crossed the threshold from hilariously bad to just bad. And then from bad to like, this is okay. And I think that's honestly where we're at right now.

22:05We're at, this is okay. It's not amazing motion design work. But if you look at the trajectory of it, it's just going to get better. And so I'm very excited for the era of a couple of models in the future, where hopefully we have models that are really, really good at this stuff. As for whether we train directly for this, I'm actually not sure. But there's certainly been a lot of examples where it's just as these models get generally only more capable, capabilities fall out that you didn't really even expect. And motion is one of these. Yeah, I guess it's all code at the end of the day. So better at coding than this stuff follow through.

22:36But I think you have some pretty awesome demos for the dashboard and the motion graphics stuff. Can you show us some? Yeah, I would love to show you. Yeah. So we've been on this broader push to make Cloud Artifacts really good. So we released a new version of Cloud Artifacts recently. We actually built our new Cloud Slides and Cloud Design on Cloud Artifacts. We said, hey, we're going to make this platform so powerful that we're going to be able to build our own design and slides tools on top of them. But anybody can really create anything on this platform. It's a general purpose program building platform.

23:07And one thing we found really compelling is building dashboards, right? And we've been doing this internally for like six to eight months where we don't really use off-the-shelf solutions for making dashboards anymore. We just ask Cloud to pull data from MCPs and code up whatever it is that fits the data that you want to track. So here's an example. I can't pull our real product metrics, but I got this nice, very large CSV of data about the New York City MTA daily ridership. And I plopped it into Google Sheets and I used the MCP, Google Sheets connector to turn that into a dashboard. So all I did was I gave Claude a link and I said, Hey, make a dashboard out of this.

23:46I actually don't and what I want to track here. So I just said, hey, make me a dashboard or whatever. And it did. And what it told me was, hey, these numbers are a saved copy of your sheet. So they won't update on their own. But since it lives in Google Drive, I can hook it up to read live. And so that's what I did, because that's what you would do if you were pulling from BigQuery or Amplitude or any of these other real data storage systems. So I said, yeah, hook it up to read live. It updated the artifact. And now when I open it up, I can reload it and it will reload the data right from Google side.

24:17So I could go and modify this data and it would update live. I'm not going to do it because I don't understand this data well enough. But you can see we've got numbers, we've got charts. And because it's all code, I could ask Claude to do whatever I wanted here. I could have it make custom visualizations. I could have it cut the data in really interesting ways. I could have it present stuff in nice ways. But this is just a really, really basic way of looking at the data. And you could have cloud code this yourself, but we just built a nice little out of the box tool for doing it that comes with some nice goodies like you can enter this sources mode where you can point at a particular number and see exactly where cloud got the data from.

24:57You could look at the actual code that it used to pull the MCP, which is pretty cool. In this case, it's not SQL because we're pulling from the Google site, the Google MCP. So it's actually just writing JavaScript, but it would be SQL if we were doing it for real. but it's neat because if you actually do understand how to do data science, which I'm not a huge expert in, you can go and you can actually make sure the numbers are coming from where you thought they were going to come from. And how do you trigger this dashboard? Would it just be like, hey, make a dashboard and it'll trigger automatically?

25:25I mean, the way most people do it is they say, hey, make a dashboard out of this and Claude will do it. You can also do it from the artifacts pane, I believe. So you can go in here and click new dashboard and do it that way. I actually haven't made a dashboard for my YouTube stuff and I just linked my public channel and be like, hey, can you make a dashboard? Give me some insights about my channel and also research a bunch of other channels that make similar stuff. And yeah, it was actually like super insightful. I was kind of blown. It's actually probably better if you don't specify exactly what you want because then it'll surprise you in a very pleasant way.

25:56Yeah, exactly. And what's cool is because it's all part of the same sort of cloud ecosystem. It can pull data in a ton of different ways. It can pull it from your file system. It can pull it via code. It can pull it via MCP. You could probably rig up a system where you have a scheduled task that uses computer use to go to the YouTube dashboard and pull the numbers out manually and put them in. It's pretty wild how much you can do here. I think sharing a dashboard in a meeting is fine, but I think what would be even cooler is sharing a sensor wheel of the stats, right? So how do I convert this?

26:25Of course, yeah. I mean, so we launched alongside this, we launched Cloud Motion, which is our, honestly, just a lightweight wrapper on top of the capabilities that people were already discovering in the model. We were seeing people externally make cool, sizzle reels and data visualizations and explainers using opus 5.5 and we said okay let's just give them a slightly nicer tool to teach people that you can do it and also to include some prompting and some ui to make it a little bit easier so here's an example of one that i have been that i was playing with last night this comes from the same data set and you'll see this is actually a long thread i actually did not get what i wanted the first time we can go through what i had to do to fix it but what i started with is i said hey use a motion artifact to tell a story with the data from this dashboard.

27:08I gave it the link to the dashboard. I asked it to give me a storyboard in animation form, static sketches, and also to make an artifact showing a couple different visual styles that are representative, which is kind of the workflow I was talking about earlier. And finally, I pasted in some little pieces of reference inspiration for Claude to use. And what I got was, let's see. So I got this storyboard, which was just a simple animation, totally static at this point, because I had asked Claude to do this. It doesn't take a lot of tokens to do, but it sort of showed me roughly what Claude was going to aim for.

27:42And I also asked it to give me a couple different shots at the visual style. So we had A, which was these cool, colorful lines. We had B, which was the Metro card, and C, which was the countdown. And I was just instantly like, oh, it's obviously A. Please do A. And so, you know, I went and did it. And from there, I got a bunch of different versions. And I think that the final one is this one. I'll play for you now.

28:15So it's going to visualize the data, but it's going to do it in like a kind of a neat stylized way to show you how people get around. A bunch of different segments in here. You can see how it changed since 2020. And you can see the sort of road to recovery in terms of ridership. We're doing all sorts of neat things with this line motif. like we're going to draw a bar graph and stuff like that. And I ended up in a place that I like a lot, but actually it took me quite a long time to get here. You know, like I said, I think we're at the point where like the animations are not bad, but they're not great.

28:48And they take a lot of poking and prodding to get in a good place. So I did a lot here to try to make this look good. So one thing I did, which is like a thing that is a little bit of shortcoming in the current models is Cloud will often make animations and it will have a bunch of things on screen changing at once. Because Cloud, of course, as a machine, can attend to a bunch of things happening in different places at once. It can read an enormous amount of text at once. And it hasn't quite internalized this idea that humans, we've got eyes, we're kind of slow sometimes. We can only really see one thing happening at a time.

29:19And so often I'll prompt Cloud and I'll be like, we need some time to take things in. We can only focus in one place. We can't read that fast. So when you're making an animation, you have to really think pretty hard about, if I'm putting some text on screen, do I leave enough time for people to read it before I show the next thing to happen? And are their eyes going to be in the right place to see the next thing as it happens? And that happens a little bit, but I still gave it a lot of feedback, right? At a certain point, I told Claude, hey, this feels like static slides. Add some motion. Add a single line that we're always following through.

29:48And then I actually was, I had a bunch of ideas. And I was walking to dinner and I went into voice mode. And I just sort of brain dumped a bunch of thoughts I have about how it wasn't really working. different idea where we picked a couple different colors. We have a line, start from a particular point. So I was visualizing what I wanted in my head and I was doing my best to kind of word vomit it onto the page and tell Claude what to do. Kind of worked. I had to do it a couple of times more, but then eventually we got to a point where it felt pretty good and it was ready for just fine tuning. At that point, I use the comments tool that we have that lets you add a comment at a particular point on the video, at a particular part of the canvas, Which is really nice because then I didn't have to say every time I saw a piece of text that overlapped, I just sort of clicked it and said, this text overlaps rather than having to say at 17 seconds in the top left corner, this text overlaps with this text, which takes a lot of time to do.

30:42I guess point, I can tell it what I want. And then I tell Claude, hey, I added some comments. Please address my comments. And it does. Oh, that's all. I wasn't even aware the motion thing had a comment feature. Yeah, we've got comments on every type of artifacts. A lot of them are just commenting in line on the text. but for certain fancy artifacts like this you can comment we override the comment behavior and we say oh you can comment at a particular point in the video or you can even drag a box around something and say like yeah make this blue or whatever okay do you have i mean you probably don't make motion graphics a ton but do you after you go through all this iteration process because i also have to iterate with it it has to be humane the loop it's not like one shot yes after this process do you have like build some sort of scale or something but like hey next time i make motion graphics, just rubber, all of this stuff.

31:26Yes. I have a skill that is framed towards internal anthropic sizzle reels. So it's like, if you imagine the types of videos that we put out for new features, it's kind of like trying to make a video like that. And actually the way I made it was I got a bunch of those videos in Google drive and I said, Hey, Claude, analyze these videos, look at the frames, look at how long it pauses between things and write this all down. And it did like a pretty decent job at it. And it's actually a fairly popular skill internally. and people are doing this all the time as a way to like get hype for an idea. You'll go into Slack and you'll see people who are just posting what looks like really fancy, well-produced videos.

32:02But really it's just anybody's dumb idea can be presented in that format. I don't know if that's good or bad, but it's certainly kind of fun. It's kind of like nowadays, whenever I launch a new skill or whatever online, I always make a little series of real with Cloud. So I guess internally you probably have the same thing. Like some people prototype something and just have a reveal to go with it. Exactly. It's one of the things like we would never do that. I don't know the first thing about After Effects. And so it's like, for me, I'm like, oh my God, this is a capability that I didn't ever have before that I do now.

32:29I kind of get a sense of what it felt like for people who didn't code when Cloud Code came along. I'm like, oh my God, I have this new capability. Based on this conversation, can you go update the scale, right? But the problem is then it kind of overfits the scale on whatever conversation I'm having. So do you like wait for it until you have like four or five conversations and then you talk to the update or do you have any tips around that? I think that the thing that I did for making this skill and updating this skill was I would ask Cloud to update the skill. I didn't actually read it. Maybe I should have.

32:56But I was basically like, okay, Cloud. I gave Cloud an eval where I was like, hey, think of the most boring features you can think of in Cloud AI and make sizzle reels for them. And so every time I made an edit to the skill, I would say, hey, Cloud, spawn a couple of Cloud code sessions and don't tell them anything except this skill and this prompt to make a sizzle reel for the model picker or something like that. and then I would look at the outputs and I'd be like oh I can tell that this has the right content in it or it doesn't so it was a very quick way to almost like an ad hoc eval but in terms of the model over fixating on particular things you tell it I think that will just be one of those things that gets better with the model's capabilities it's a judgment thing it's a kind of like reading the room thing one of those very fuzzy things that's hard to prompt for but eventually we'll just wake up and it will be fixed when you say you give it eval you mean like you're personally when you look at the results or you have a judge or something.

33:49Yeah. I mean, for something like this, it would be more work, I think, to try to write an eval for good motion design than to just look at a couple examples. Got it. Do you have any tips? Because I also build a bunch of evals for my scales, and a lot of them are like yes-no evals. Because my understanding is that that's better than having a score out of five or something because Cloud doesn't really know what four versus five is. Do you have any tips for that kind of stuff? Yeah, I mean, yeah, definitely agree on yes or no is much better than one through five. Rubrics, in the same way that you would evaluate an employee or somebody in an interview or something, where if you want to have a one through five score, you have to say, here is what five looks like.

34:25Here is what four looks like. Here is an example of three. Here's what's missing if you're a two. And you can do that. I mean, it really just boils down to being a good communicator. And the same things that you would do if you were a member of an HR department and you were trying to come up with a really good employee performance review system. Just think really hard and deliberately about that and communicate it really clearly. As for design evals, I mean, honestly, like it's just so hard to do design evals. We've tried a lot. And the problem is if you assume that the model doesn't have amazing taste, then you can't use the model to subjectively grade whether something is good design quality.

35:04And so there's not that much you can do in terms of evals besides the very basic stuff, like does this text overlap? And did we assume, can we ensure that if you give it 10 prompts, They're not all the same font and all the same color. So stuff like that we eval, but it's pretty hard to get a little more in the weeds about what is good design, what is good motion. Got it. So you just need good little human eyes on it. Yeah, I think that's the big thing that we do a lot is you read the transcripts, look at the outputs, and you iterate. And I'm glad you brought up the taste word because all the designers are talking about taste and everyone's talking about taste.

35:37What are your hot takes on this? Like, do you think, what even is taste? And do you think AI will eventually get good at taste? That, you know, tastes not like human anymore? I have no idea. I feel like it's one of those philosophy things. Like, it's up there with what is art, right? What is taste? I think you could make the argument that taste is something that is innately human in the same way that you could probably make the argument that art is something that is innately human because you're trying to say something about the human condition. I certainly think that today a common failure mode with these AI tools is you ask them to produce something and they produce it and you're like, oh, that's slop.

36:11And I think that will probably go down. The models will get better at producing more diverse output. I walk around New York City today and it's the number of chat GPT generated deli ads for bagels. It's like you can see them everywhere. It's the same font, exact same layout, the same like crazy stuff. You've seen it. I imagine the models will get a lot better at like the diversity of outputs. I like to think there is something intangible and uniquely human about this that resonates with us. When somebody has something they really want to say and they really want to communicate and they're really proud of something, that will always come across in some way.

36:42But it will probably be different how we think about taste today, just in terms of visual polish and refinement. Yeah, I think taste is going to become like a personal thing. I think the models will get good at overall taste, but maybe your taste is different from my taste of what good is, you know? Exactly. And it'll be a lot more about authenticity, probably. And maybe what that looks like is something that feels a little bit handwritten in the same way that Claude is pretty good at writing. And I honestly, whenever I get a typo now in an email, I kind of value it more because it's like proof that there was sort of like proof of work behind it.

37:12I know somebody actually took the time to type this out. And I'm sure we will find the visual equivalents of that for animations and designs and stuff like that. And we will continue to value them a lot. Yeah, you got to type all your updates with lowercase and no punctuation. that is a real human doing it. All right, just a couple, one last question about how you like to work with Cloud. Now you have Cloud Code, you have co-work. Like some of these engineers are obviously using one orchestrator to orchestrate, I don't know, a hundred different threads. But I'm really curious how a designer does it.

37:43And also, do you use Duet more or do you still use regular Cloud? Yeah, I use regular Cloud only because it has access to our code base. I do a lot of design in our code base. My hot take about Duet is that I hope that our whole product just works like that at some point. For my biggest design projects, I will typically actually just have one thread in Cloud Code and I'll trust Cloud Code to spin up sub-agents to do the various tasks. So for example, when we were designing the new version of Cloud Design, I had a Cloud Code thread with a bunch of design artifacts. And I think at the end, I wound up with like 36 different artifacts in that one chat, which was not amazing, but it was nice because they all shared context and I never had to worry about did they know what they were doing.

38:25And I have found that Cloud has gotten very, very good at handling very long threads really well, internally figuring out how to manage its own memory and how to manage its own context and its own subagents. And so I typically will work with either a tiny thread for a one-off thing or one big long one for a month. Yeah, that's all I do too. I kind of like to think about each thread as having its own job or its own project. and I probably can actively manage more than like five threads at once. It's too much. I mean, we do use Cloud Tag like a ton internally. I'm sure people have told you this before, but like we are obsessed with Cloud Tag and no exception there.

39:03And what's amazing about Cloud Tag is it's just the same space that you're already checking, right? It's like I already have to check the messages from other people. And so I'm also going to go and check my clods in the same space. And I just treat Cloud Tag like one guy. I don't treat it like a bunch of different threads. I don't worry about context at all. I just assume that Cloud Tag is going to have the memory from every time we've interacted and every time it's interacted with anybody else in the company. Obviously it doesn't literally have that, but it's close enough that I can kind of just be like, hey, Cloud, you know the thing we worked on last week?

39:32Well, we got to make these changes to it. And by the way, where is it? Can you pull it up? And it tends to work. Yeah, I think Cloud Tag is probably the best human and agent, multiple human, multiple agent you can experience, right? Yeah. That makes sense. For sure. Well, dude, I have to ask you timely. It looks like Arc is getting deprecated. So you were an original designer, right? So how do you feel? I think as like, you know, designers, PMs, sometimes our stuff that we're really proud of gets deprecated or something. How do you feel about that? How do you think about that? Yeah. I mean, I think the lasting legacy of the browser company is, first of all, D is pretty good.

40:06And secondarily, I think that we've kicked off this wave of innovation in browsers. The thing we always said at the beginning of the company was, I would show people a screenshot of Chrome from 2008 when it launched. And I would be like, what's different about this than the Chrome you have now? And the answer was like, nothing. It's a different color. And that's it. And now it's like, Chrome is actually way different than it was six years ago, which is crazy. And again, I'm not going to say the browser company should take credit for it. But I am very happy that we live in a world now of a ton of browser innovation.

40:37Tons of people are making their own versions of browsers. and I'm excited to see what we get soon. Yeah, that's probably like a really good outcome, right? Even though your product might be duplicated, but you introduce a new norm that mainstream product has now. So I think that's amazing. Yeah, every so often I ask Cloud to design something and it will be, for whatever reason, it will be like, oh, here's an Arc style sidebar. And I'm like, wow, that's crazy that you call it an Arc style sidebar now. So I thought that's going to impact. Cloud does that? Sometimes. That's so funny. That's so funny.

41:08Claude also keeps saying like oh it's going to take like two days to do this and then it does in like five minutes. I know I feel kind of bad I'm like Claude's gonna be like this is gonna be like six weeks work and I'm like I hate to break it to you Claude but we're gonna do this together in like the next 20 minutes so you better get cracking. I don't know I don't know why that is some sort of quirk in the model but it's pretty good to hear it say that. Yeah it's a good meme yeah cool well this is an awesome conversation thanks for showing me everything and I hope this episode inspires more people or designers or not to just create more stuff and that's high craft.

41:41Where can people find you online? Mostly my website. It's nateparrot.com. Parrot like the bird, but with two t's on the end. I'm also on Twitter under the same name and those are my main vices. Cool. Well, keep cooking, sir. I'm really excited to see where you launch or design next. Thank you. Wonderful chatting with you.

From the publisher

Copy my best AI workflows to save time and automate busywork: https://www.behindthecraft.com/?utm_source=spotify&utm_medium=podcast


Nate is the designer at Anthropic behind Claude Design, Artifacts, and more. He gave me exclusive demos of how he designs new Claude features, interactive dashboards, and motion videos from his phone. Nate also shared 5 great tips to avoid AI slop when creating these artifacts. This episode is a must-watch if you want to see how Anthropic designers actually work.


Nate and I talked about:


(00:00) Nate’s 5 tips to stop Claude’s AI slop

(05:31) Demo: Designing a Duet feature from a phone note

(11:01) How Anthropic designers prototype in days

(13:06) Why Claude Design launched as a separate product

(18:55) The 4 tells that make a design look AI-made

(22:36) Demo: Turn any spreadsheet into a live dashboard

(26:17) Demo: Turning the dashboard into a Claude Motion video

(31:11) How to test a design skill when AI can’t judge taste

(35:32) Will AI ever have good taste?


Thanks to our sponsors:


Riverside: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang

Granola: AI meeting notes that don’t suck https://granola.ai/peter


📌 Get the takeaways: https://creatoreconomy.so/p/how-to-fix-ai-slop-with-claude-design-nate-parrott


Where to find Nate:


X: https://x.com/nateparrott

Website: https://nateparrott.com


Subscribe to this channel - more interviews coming soon!

More from Behind the Craft

All episodes