How the engineer behind Claude Cowork actually uses Claude | Felix Rieseberg (Anthropic)

AI is used poorly if it just needs to move the mouse cursor for you. I want AI to do a bunch of annoying things in the background to free you up for your creative energy. The biggest gap that I see, it's not the capabilities of the tools. It is literally people being able to understand that almost any problem can go into these tools. I have built this little thing, which is just like a teeny tiny claw on a little stick. And this stick has Wi-Fi, has Bluetooth. I want my little claw to live on this thing, and I want it to cheer me on every single time I do a good job and also every single time I need to approve something the claw is doing. I want it to be on this big button that is out here. Claude has built all of that in one shot. I needed to correct absolutely nothing. My nine-year-old is a daily active Claude user, but he's decided he's really into cybersecurity. We're like very hacker adjacent here. And so he is truly Claude on one thing, the terminal, other, he's like, mom, do you know that your device ID is this? A truly magical thing is happening with kids because they've never learned what not to ask for. And I think our generation, we're very used to things just not working. So we've been living in this mind prison for 20 minutes. Welcome back to How I AI. I'm Claire Vaux, product leader and AI obsessive here on a mission to help you build better with these new tools. Today, we have Felix Riesberg, who leads some of our favorite Claude products over at Anthropic. He's going to walk us through how you can solve almost any problem with Claude and show us how for just $20, you can build your own hardware Claude buddy that you can smash a button when you need to approve things. Let's get to it. This episode is brought to you by Magic Patterns. Today's engineers use Cursor and Claude code to ship features in hours that used to take weeks. If you're a designer or PM, you've probably felt a shift too. The pressure to move faster, validate sooner, and keep up with a team that's operating at a completely different speed. You've already tried AI prototyping tools to close that gap. But if your prototypes don't look like your actual product, it doesn't matter how fast you can build. You still end up redrawing it by hand. Magic Patterns takes your product team from idea to production and works from your real design system. When you build a prototype, what you get back actually looks like your product. You'll validate faster, get alignment sooner, and when it's time to build, engineers can connect your prototype to Cursor or Claude code with the Magic Patterns MCP to pick up where you left off. Your eng team has their AI advantage. Make Magic Patterns yours. Try it today at magicpatterns.com slash howiAI. Felix, welcome to How I AI. Thanks for joining us. Hi, Claire. How are you? I am great. We were joking before we got started that I have referred to 2026 as the year of our Claude, 2026. That is the era we are all living in. And you are in charge of a lot of the experience that we every day get to get to play with and work with. So tell me, how much of Claude do you get to product? Normally when I introduce myself, I say I work on Claude co-work, I work on Claude code, then I work on Claude for Chrome, then I work on the Claude desktop applications for MacOS and Windows because obviously it's a team effort. But I'm currently the engineering lead for those products. For those products. And what's really funny is, I'm going to give you a little bit of a hard time, but every time I talk to someone about Claude and how they're using it, I say, which of the tabs is your favorite? I know, I know. Yeah. Yeah. Which one is your favorite? The one I use the most in the desktop app is co-work. I'm really trying my best to figure out co-work because I work with a lot of folks who are just uncomfortable in the Claude code. What is a terminal? Interface? Even that third tab stresses them out a little bit. So I've been spending a lot of time banging my head against co-work, scheduled tasks, live artifacts, all that stuff. So that's the one I'm living in. I think we live in this super interesting time right now where I often explain it to my friends and also to people who work here as the time right before we all came up with the glass pebble as the correct shape for a phone. I think we currently live in this super interesting time where some people are like, should your phone be shaped like a taco or should it be shaped like a lipstick? Should it have the keyboard underneath the screen or right next to the screen or no keyboard? I think that's the time we currently live in, and I think you can see that in a lot of products. I think that's also why you probably see more entry points into Claude because I myself, as an individual, I think it's a super interesting year because we're playing with all of these different form factors, and I have not yet gained the confidence to say, throw them all away. This is the one. I'm perfectly happy if people sort of pick their favorites, but I do understand that simply giving people a choice, right, like the same way that a restaurant does, creates a little bit of like extra work that we put on the user, and it's true that if I fast forward to like, what is the end state for this, I'm probably not going to make people pick anymore, right? I'm just going to, whatever you want to do, here's the one place where you do that, but right now it helps us quite a bit to know, are you just showing up to get some quick answers? Are you showing up to like do deep, meaningful work? Or are you showing up to do engineering work? Like those are sort of the three buckets we put you in right now, and knowing that ahead of time makes it a lot easier for me to give you an experience that is really, really good depending on what you came for. Yeah, I like that because a lot of what I tell folks is we are really pre-convergence on AI tooling just generally and AI experiences. And so my answer to, you know, how do you want to, how do you want to, it's like, why not all of them? Just depends on the day, depends on the day if I want my phone to be a taco or a glass pebble. And it honestly depends on what mindset I'm in. Well, we're going to talk a lot about product at the end, but to start, this is how I AI and we want to know how you Claude. And so we were talking before we got started and I said, you know, we've done a Claude co-work 101 episode with JJ. We've done a lot of Claude code in the terminal episodes, but we want to know how, how you use it because you probably know the edges and secret features and some workflows that maybe folks wouldn't think of. So I'd love for you to walk me through one of your favorite Claude, maybe co-work workflows that you're using these days. You'd think that I would be much better at using the tools than other people, right? Because like my entire job is to like build them, work on them. And an interesting thing is happening where whenever people show me something like super impressive in how they, how they use Claude, like any of the product services, it is pretty rare that I'm impressed by something the model is capable of. Like I sort of know the model capabilities. We have a pretty good understanding of what the model is good at, what it's bad at. And usually what impresses me is like the creativity in structuring your work and coming up with an idea. That is usually what impresses me quite a bit. But my favorite example sort of like changed week over week. It kind of depends on like what I'm, what I'm currently working on. Claire, you and I, like right before we started the interview, like briefly talked about our kids and like how, how we both recently had kids. And like, if you'd asked me three months ago, I would have said, oh, like, just putting all my medical documents into a folder and like constantly annoying Claude with all the questions I didn't have time for in the hospital. It's like an excellent use case. But one that I've one that I've quite enjoyed recently was, I'm about to move. I'm about to move into a new place and my realtor just gave me the little like marketing floor plan that you get from, from a real estate company. And it didn't have any units in it like whatsoever. And this is like the kind of thing where I turned to co-work and what I've done is I have a folder and I like made, made a very similar, sorry. I made like a similar fake folder that doesn't contain all the same like secret stuff, but it contains fake documents about like an artificial house and an artificial house I'm like picking. I'm just going to select this here. I'm going to say always allow. It's the demo folder. And the demo folder contains all the disclosure, disclosures you would get. It's a floor plan. It contains like information about the mortgage, any kind of like records that I could find about the house. And I remember saying something along the lines of, in this folder, you can find a floor plan. Can you figure out what the units for that floor plan are and maybe make me a new one with units, right? And you can fire this off. This doesn't actually need that smart of a model. I've actually, I actually did that with Sonnet. So you can like just fire that off and it's going to go off and do, do that work. This is a fake house. So it's not going to do nearly as interesting of a job, The furniture. And then I went another step up. I was like, again, what am I doing? Why do I need to tell Claude what kind of furniture I have? Just like, you figure out what furniture I have. And like applying this like step up in abstraction of the task and doing that over and over and over again. And then like spending all of my energy on the piece that is maybe like more fun and more creative, which is like steering Claude on what the house planner should be like or like adding stupid little features, like I want to be able to walk through the house myself in 3D like it's a video game. Those are the things that I spend my time on. And I spend my time less on like figuring out what was my furniture, when did I buy it, who did I buy it from? Because I don't know for you, but for me, certainly, I did have this problem that I look at my house and I'm like, I have no idea what bought all of this stuff. Wait till your kid starts bringing more and more stuff into the house. And I do a very similar thing. I've been spending a lot of time with executives trying to figure out how to scaffold their own operating system with AI, a lot of times with Claude co-work at the center of it. And my rule, I call this like the anti-to-do list, which is anytime you're like doing something that's tedious, I just, I'm like, I smack their hands. I'm like, why are you doing, you don't do that. Your hands are not allowed to touch this. You need to go ask. Again, go that abstraction layer up. How could Claude do this? And then how could I never do this again? I think is the next question, which is if you want to keep this furniture or household inventory going forever, the next abstraction would be like, what's the system that in a year or two years, if I have to ask this question again, it it's already there and ready for me to go. And so that's the other piece that I think people kind of miss is going that next layer up, which is like, how will I never have to do this again? Yeah. Yeah. This is really good. This is really, really good. Like what you're saying makes a lot of sense to me because like I also frequently, I frequently then take the same step to what you're saying, okay, like how do I keep this going? How do I make sure I can continue using this and like the new data makes it and that continues to be helpful in my life on an ongoing basis, not just a point in time. And like a thing that happened to me recently is a lot more people have figured out that they can just like message me on Twitter to like solve that problem, which is nice. But it created this problem for me where I now suddenly have to keep track of my problems. And I did briefly like, sorry, not keep track of my problems, but keep track of my promises, the promises I make to people, right? Where I'm like, send me your logs. I'll look at them. I promise you. I'm going to figure out whatever went wrong. And that too is a problem to just give Claude. I was like, Dear Claude, you can read all my messages. I want you to keep track of all my promises. I also want you to figure out how you can do that without like rereading all the messages every single time I talk to them. And all of these things Claude is really good at. I have never actually looked at like Claude's little database. I think it like created both a SQL lite database as well as like a mountain of text files to keep track of all the promises I made to people. But it's quite good occasionally reminding me of like, Hey, just so you know, two weeks ago, you said you would do X. It is time to do X. Please go do X. What I like about what you're saying too is while you might be curious about how these things are built, you also aren't getting in your way being like, how do I build the perfect 3D renderer of my house? Or like, oh my God, I cannot believe this is text files. This is so ridiculous. Like as long as it solves the problem, you can sort of move on, move on with your life because these are, you know, little micro apps. They're just for you. They don't have to be architecturally, you know, perfect or the code that you would write necessarily, although they're probably pretty good. And so you really get to like how to solve the problem versus, or that the problem is solved versus how you've solved it being the most important thing. Yeah. Yeah. I think that's totally true. Like I increasingly started stepping away from, from actually like supervising Claude too closely and really only judging it on its impact, right? It's like something that I think for people in the workforce, we often aspire to is that we're really only being judged by like the actual output. And I think as a control freak who like, you know, for a long time really needed to like read every single line that Claude writes anywhere. It was a little difficult for me to give up that level of control, but I'm, I'm getting increasingly comfortable consulting Claude with things where maybe I'm not the ultimate authority myself, right? Like there's a lot of things I'm good at. There's an even larger amount of things that I don't know and I'm not very good at. And I think a lot of us are making, making this experience for the first time with getting just a little bit more context on medical issues. This is like, this is like a thing that I think is very universal when I talk to my friends. Like when, when they have medical issues that don't fully understand, at least not completely. And they just have like some questions about the physician's advice. They ask it both to the physician as they should, but they all of them without fail will also like ask an AI model of their choice. Hey, can you just explain this to me a little better? Like I have the following 10 questions about this, right? Like, and in a way there, AI has replaced, you know, the usual Google endpoint for I have the following question about life on the Internet. And knowing, knowing that I'm quite comfortable with, with using these tools in a domain where maybe I'm not the ultimate authority has also made it easier for me to not double checking every single thing it does because to your point, like if the end result is good for like, I have no, like I am obsessed with software. I've not read a single line of this financial planner. It's just for me. I'm just going to play within myself. It is really only for me and my wife to like design our house. It will get thrown away in a month, but in this moment, it is very, very fun. And I don't need to optimize it for like a million users. This episode is brought to you by Guru, the AI layer of truth for your company's knowledge. Here is the problem. Your AI is only as good as the information you feed it. Most companies are getting confident, but wrong answers from AI because their underlying knowledge is outdated, incomplete, or just plain incorrect. Bad information doesn't just slow you down. It costs you money. It puts you at risk. Guru solves this by adding a verification layer between your company's knowledge and AI tools. Instead of just hoping your AI gets it right, Guru automatically scores content for accuracy, flags outdated information, and ensures your team gets trustworthy answers every time. It works with the tools you already use, so you don't have to change how you work. Thousands of companies trust Guru to keep their AI accurate and compliant. Ready to stop playing Russian roulette with your company's knowledge? Visit getguru.com to learn more. I want to go back to something that you said, which is, you know, you built this promise tracker, which I think is really interesting. And that made me think about a new, a new feature in Claude that I've been using and playing with a little bit, which is live artifacts. And I think this idea, you know, actually, I'm not going to, I'm not going to pitch live artifacts on your behalf. I'm going to let you explain what it is, and then maybe you can give us an example of where live artifacts could be a helpful tool because, because it reminds me a little bit of what you're building with your furniture builder. So, so tell me a little bit about that and Claude. Yeah, yeah. Here, let me like share my screen again. And I'll give a little bit of a quick spiel of like, what are artifacts, what are live artifacts? Why do we have them? What's the point? So, generally speaking, what do we call an artifact, right? Like I can say, dear Claude, please make me a beautifully formatted page with two poems on it. Format it nicely. Beautiful. I like, I like the letter form. Yeah. The lack of creativity here in the prompt is like, it's like pretty rough, but... It's beautiful and nice. What Claude is now going to do, obviously, is Claude is going to figure out, okay, Felix wants me to write some poems and then those should add up, those should end up on a page. And this page is what we call an artifact. We, within the Claude world, an artifact is like some kind of file output from what you've been working on. The furniture planner I just showed you is a single file. That is an artifact. Any kind of PDF you will make is an artifact. Spreadsheet, presentation, all of those are artifacts. And here we go. There are my poems over here on the right. This is a classic artifact. And artifacts become really powerful if you, if you use them in a way that they help with your life, right? Like the The widgets I came up with are on the spot. This is like clay, right? This can be anything you want it to be, which is very open-ended, but therefore so powerful. People frequently use this to get a little report about what they want to do in their day, right? Like just prep me for my day, what are the meetings I need to like think about? But again, the thing that is maybe most powerful here is going one abstraction layer up. So I actually don't believe strongly in morning reports that just summarize what kind of meetings you have today. That is not all that impressive. I can just look at my calendar. What is really impressive is saying, look at all of my meetings I have for the day. Go figure out who I'm meeting with. Catch me up on like the recent conversations we've had. What were the themes there? What were the problems? If it's a coworker, go on Slack, figure out what kind of stuff they've been working on in the last week and what stuff they might want to talk to you about today. And then what you need to know about that concept. You can just like keep playing this like graph bounce where you go and pull in more and more information to get you prepared even better for whatever you like for work is. And that's where these things get really powerful. And this is the kind of data you can pull in live and have Claude crunch on the go. Well, and I want to call out for folks a couple of things that they might not notice that I think are really unique about these Claude live artifacts, which is one, the killer feature here is this little refresh button up in the top, which is you can constantly pull in and reload data. And so this isn't, this is your day today. Every time you come to this, it's going to refresh your data and pull the latest in. Then the second thing, and correct me if I'm wrong, is it just is reusing the auth from your connectors. And so there's like no magic. I have to put an API key in and where am I going to store it? It's if you go into settings and go into connectors and you connect your stuff, you can use any of those kind of connectors to scaffold out one of these live artifacts. And like, look at this list of stuff. There's so many things that you can just one click, add in, sign in through OAuth or whatever, and then you can build a live artifact in Claude. Without having to like ask for an API key. Exactly. And it can be, you know, it's not just the ones we have in our, in our little connector store. You can also connect your own. It is quite powerful. And the thing that I enjoy about this the most, and this is like probably fairly unique to myself, but the thing that I enjoy about this the most is the amount of creativity you can pull in and how you can really shape this in your image, right? Like we could go in and say, I want this to look like it is software made in the early 2000s. Oh, this is my. 200, maybe. 200 is too early. No, let's see what it does. I mean, honestly, this is Claude. Claude is probably smart enough to know that I meant the 2000s. Well, this is my go-to Claude design use case, which is I did an episode on Claude design earlier this month and I redesigned Lenny's newsletter in like 1999 GeoCity style. It is incredible. We'll put the link in the show notes. It did an impeccable job at this. And what I told people is it wasn't even the design that was really impressive to me. It was the copy was so. Oh, was it like, yeah, was it like GeoCities era? It was GeoCities era copy. It was 11 out of 10. Exceptional copy. Beautiful. I want to see if it's going to do 2000s or 200s. 200s would be way funnier. We can probably look at it already, see like it's really refreshing thing is 2000s. Like Claude is smart enough. It knows too much. It knows too much. And also, I mean, this is obviously like a bit of a demo account, but if you, if I showed you my real account, Claude is like quite used to the fact that I constantly misspell things and like talk to it in a slightly different style. I'm like, I'm like eager to see how this is going to go. I'm in general quite excited about the amount of creativity you express and especially the amount of creativity people can express who are not software developers, right? Like this used to be this entire domain where you were maybe a designer working at a software company and like, maybe then software developers would execute your vision or you like picked a medium where it's a little bit easier. You can like low code things. But I think what is really exciting to me is that we can now pull in people from all these different, from all these different like expertise areas and they can all build these artifacts, like either a presentation or be like a straight up tool sort of in the vision they have. A couple of weeks ago, I read somewhere that Margaret Atwood wrote about using Claude and the immediate thing that popped in my head was, I want to see software written by Margaret Atwood. Like I want to see what like tools look like that she has made for like writing her own books. That would be like so interesting and exciting to me. Okay, Margaret, if you want to come on How I AI, I would love, love to see your Claude setup. And I want to reflect on this while this is, we're going to make this happen. While this is loading, I do want to call out because I do think there's a gap. People listen to and watch How I AI along a broad spectrum of technical capabilities. So we have the most AI-pilled, I float through the tokens, intense software engineering folks that are really pushing the edges of what these tools can do. And then we have very, very early users who are just coming to this in sort of a naive growth mindset way and saying, I need to solve a problem. The biggest gap that I see, which we were again talking about before we started recording, is it's not the capabilities of the tools. It is literally people being able to understand that almost any problem can go into these tools and what like a workflow would be. And this is why we've done this podcast. And on the ChatPRD website, we have, I think it's now close to 200 AI workflows that we've listed from this podcast. That's just like, if you're trying to solve a personal problem with Claude code, here are the 15 ideas that we've seen on the podcast. And I do still think there needs to be this education and muscle memory change of you can build yourself a daily dashboard. You can build yourself a newborn sleep tracker. Hey, you're moving. Why not throw it all into a folder and then make your life easier as you're trying to close on the house. And I think people still, the gap between technology and use case is still pretty, pretty broad. And so, you know, my advice is like, if you're futzing with something on your computer, ask Claude to do it. Any other, any other tips you've seen on how to like discover use cases? Yes, but I think before I even give you an answer, I just want to like underline how correct I think you are. Many years ago, I spent a decent amount of my time, like I spent five years at a company called Slack, which I think now is like fairly well known as like this chat tool that a lot of people use. And people were sort of like using Slack for different reasons, but they were all more about the way they organize their work with other people rather than it was about the tool itself. So Slack by itself, if you used it the exact same way you would maybe communicate with emails in a silo, was not all that useful, right? Like Slack becomes really powerful if you open it up to the idea of, what if you communicate more transparently and people can see what's going on? And we like, sort of instead of having email threads, we have a channel and like the channel is open to anyone that wants to see what's going on, right? Like it's this transformation of work and the transformation of how do you actually organize communication with other people? And AI is like in a similar vein of, it like opens up so many opportunities that to harness them all, you probably have to change a little bit how you work. Like if you really want to harness them. And yeah, I think you're right that like one of the biggest problems we have as an industry overall is not even like training the models and making the models very smart and making them very powerful or giving them the tools that allow them to turn that intelligence into like output. The biggest hurdle we have is how do we design the human model interface? How do we make this very easy for humans to harness? And right now I think the way it shows up is that humans are sharing these like tips on how to use this resource, right? I think the tip that you gave was a good one. It was like constantly, whenever you do something that you find annoying and you're not enjoying and it doesn't feel creative and it doesn't feel like something where like you're, you know, it doesn't feel like the core of what you really want to do right now. That is a good time to pause for a second and like wonder, is there a way I can use AI to do this? And then of course, and this is probably not surprising, but of course, like many people have a decent amount of success just like asking Claude, right? Just ask Claude, how would you help me with this task? And usually comes up with a pretty good idea. I other people have like built similar things because I open source my code so people can play with it. But I've also since seen that this is just like the tip of the iceberg. There's like so many more cool devices out there. This, by the way, so this thing is $19. It has Wi-Fi. It has Bluetooth. I already mentioned this like many times, but I can't get over, I can't get over the idea. We've been talking a lot about like how we optimize our lives sort of as it happens on a computer, right? Like because that's where we live. But I keep having this idea that it can just like unleash Claude onto like all these little hardware devices and also improve my improve for me and just make like more delightful, make a little cuter. I did a very similar thing in a recent episode. I bought, it's like, it's already preloaded with this proprietary software, but it's like a tiny retro computer, very Windows 2000. And I had to hack into that thing. I was like Bluetooth sniffing between the like iPhone app and the thing. I was like dumping all the logs in and being like backwards engineer the encoding of the files into this thing. And I finally got it and made myself a little CLI tool where I can type in a But see, that is so cool. That is so cool that we can do that just now. Like I, this thing also communicates over Bluetooth. I learned a lot about the Bluetooth protocol as I was building this just by osmosis. Do you wanna, do you wanna see it? Do you wanna see this? Yeah, I do. I do wanna see it. Okay, so this is now also like partially live in actual Claude, so you can build something very similar if you'd like. Buy any IoT device of your choice. I'm not gonna make a recommendation, but anything that you find on the internet will probably work as long as it has Bluetooth. But you go into help, troubleshooting, you enable developer mode. Developer mode allows access to developer tools and features. We give you that warning because it does require that you sort of know a little bit of what you're doing. It's going to restart the application. Here we are. And then you have this new developer menu where we hide all of our developer features. This is for people who work with MCPs or also for people who want to build their own little hardware devices. And in here is now a new open hardware buddy. This thing, you will notice that it looks a lot like the one I've actually purchased, but it doesn't have to be the same one. I think I made that point clear. Follow your heart. Be creative. And you can hit connect and it's going to like scan for this little thing for like a few seconds, like find it on Bluetooth. And it's found it. There we go. So now it's connected. It gives me a passcode. Okay, cool. So we're now connected. And Clara, you tell me if you can see anything. So maybe what we're going to do is we're going to ask for like anything that would require any kind of like permission. I'm just going to say, just make a hello world.txt and write it to disk. And the point of this is to just fire something off that will require a permission. Okay, so you can now see. I heard it. Isn't that cute? Isn't that very cute? And if I approve this here, then it will be approved. That's how that works. And now Claude is off doing something. I love it. We saved the best for last, people. If you have made it this long in our episode, we have saved the best for last. Felix, man, I really like it. And I'm like you. I'm like, I'm a perfectly serviceable software engineer, and I have no idea what I'm doing with the Bluetooth protocol and with hardware. And it's just been a thing that I've stayed so far away from. And you'll appreciate this, I think, more is as my kids get older too, I don't want them interacting with like the full computer thing, but I would love to have little, little things for them to interact with. I think like being able to scope kind of like micro devices to a use case, just like we're getting this personalized software is gonna be really fun. Yeah, and like, one of the pieces that's been maybe most joyful about my job is that people frequently send me like little videos of the kinds of things that kids are doing with Claude. And the creativity they have is like beautiful, right? Like you can like just take a photo of like a little animal your kid has drawn, give it to Claude and be like, I need a jump and run. I need a flappy bird, but instead of the bird, it needs to be this little drawn bird over there. And then you put your kid in front of the computer and like truly, truly magical things happen because I think I'm still wrapping my head around this, Claire, but a truly magical thing is happening with kids because they've never learned what not to ask for to them. The computer is just like, can do anything. And I think to like sort of our generation, we're very used to like things just not working, right? We're like used to like, oh, you can't do that. So like, so we've been like living in this mind prison for like 20 years. This is very on brand, but my nine-year-old is a daily active Claude user, but he, he's decided he's really into cybersecurity. We're like very hacker adjacent here. And so he's like truly in like Claude on one thing, the terminal other. He's like, mom, do you know that your device ID is this? I'm like, yeah, babe, I do know that a good job, a good job. Yeah. I think it's totally fascinating. Well, Felix, I know we have just been having so much fun this episode. I think people are gonna get so much out of it. Just to reverse for folks. We talked about Claude co-work, being able to build you artifacts, both sort of like standard crazy 3D printed or 3D renderings of your floor plan and other things that you can just stand up to solve little problems in your life all the way to life artifacts that can pull from your live data and really give you a rich personal experience with your data. We touched on a lot of topics, including email as a source of truth data, what we're gonna do with all this latency and how to pick the smart enough model, which is to just say, you know, you know when you need Opus and you know when you need Sonnet, but we both think that Sonnet's pretty good at most things. And then you showed us how we can hack into hardware for the all in price of 20 bucks, including shipping and build little, little buddies for ourselves. I'm gonna get you out of here. One quick lightning round question. Again, we did this at the midpoint, so maybe you can just answer really quickly, but when Claude is not listening, we've, we've talked about how to set it off on a good path, but when it's really just making mistakes, instead of making no mistakes, it's making all the mistakes. What is your prompting strategy? I can't imagine you yell. You don't seem like a yeller, but what, what do you do? Yeah, I do not yell. I think, I think it is, it is useful for me to know when people curse at Claude. This is like, this is like an interesting thing for me to know as a product builder. But my strategy for when Claude doesn't like really, really understand what's going on or like Claude seems to have gone off the rails. I usually try to debug my workflow straight up by asking, okay, here's what I actually expected. Can you maybe walk me through at what point you expect something different or like how we can maybe prevent this in the future? It doesn't always work. Claude is AI and AI can make mistakes. However, however, that being said, I, I have frequently found ways to improve the data source or to improve the amount of noise that is in the data, more like with Claude to come up with mechanisms. Okay, how do we, how do we like, do, do we set up some kind of dry run? I usually end up at a place where the learning I take away is not, oh, Claude can't do this. I'm not gonna have Claude do it. Usually the learning I take away is I can maybe change a little bit the harness, the prompt, I can change something and like how we set this up at the beginning. And I think after, after the time we had together now, there's no surprise to you, but like usually my answer to that is like to also involve Claude and say, Claude, what do you think we can do here? Claude is a good friend. And if people need, if people need to vent and people need to like externalize their frustrations, Claude can take it. You can curse at Claude. You can even do it. Maybe just smash the button, like have an angry button instead of saying the thing on our little hardware buddy. You can have like the mad button and the happy button. You could just, I'm not pleased, Claude. Smash, smash, smash. Yeah. Yeah. And like there, there is a I'm not pleased button that actually also ends up on my desk that I do want people to like use. So every single time Claude does make a mistake, we do have the little, like in many of our product surfaces, this is a cross Claude. You will find the little thumbs up, thumbs down buttons. Those are super useful to us. We use those in training. We use them in like improving, improving not just the model, but also the products. So maybe actually the thing I'm gonna say is like, if Claude hasn't done the thing that you wanted it to