Claude Cowork Tutorial from Cowork's Design Lead (40 Min) | Jenny Wen

They're creating entire features in days, not weeks these days. I think about the specs that we used to make 6 months to a year ago and, you know, be like milestone 1, we have these P0s, we have these P1s. We still do that prioritization, but it's usually just like a few bullet points. My secret is that I actually use Cowork for most things now. I start every Monday morning at 10 AM with this presentation and with like 3 different product ideas that I can use and kick off the week for. The UI looks so different just like 4 or 5 weeks ago, and now it's like we're constantly learning like what's working and isn't working. If you feel like the ground is shifting beneath your feet, it's because it is.

Hey, everyone. I'm super excited today to welcome Jenny, design lead at Anthropic. Jenny is going to show us how she uses Cloud Cowork and Cloud Code to design and ship products and give us the inside story of Cowork and also maybe what's next for the product. So welcome, Jenny.

Thanks, Peter. So good to be here.

Yeah. All right. So I'm just really curious, like, What is a typical day like at work for you? Like, what tasks take up most of your time?

I'm like, is there a typical day? Like, I don't know if there is a typical, typical day. But a lot of the things that I spend time doing is it's like to get product out the door, right? Like, so, but I think that maybe looks different from what it looked like a year or 2 ago. And a lot of that is, a lot of it is just actually just like jamming with engineers and product people and whatnot in like a less formal way. It's often, looking at a prototype together and then like pointing out and thinking about the ways that it could evolve. It's sometimes just like talking about the behavior of things. And then sometimes it's me like implementing stuff. I think there's also still a big portion of the time that is like, you know, me designing stuff, prototyping and stuff like that. But it feels like a lot of the way that design works these days is kind of loose. Like it is in these— it sounds so like Like non-tangible. It's like, oh, it's a conversation, or it's just like jamming with people. But that is kind of what I find a lot of the work to do, because to be, because it's less, you know, I spend less of my time just focused on one single project. But I feel like I'm like consulting on like, you know, 5 or 6 different projects at any given time.

Interesting. Okay, so basically, you have a bunch of prototypes that generate through like Claude Code or something, and then you just kind of jam with the engineers, give some feedback. And they just prompt AI to improve it, right? Is that how it works?

Basically. And they're not even often prototypes like that. They're working prototypes. They're things that are in our internal builds and instances of Claude or Cowork. And usually, I've spent some time using the feature, pushing the feature, seeing what it's capable of, and forming opinions. And the next step on iterating with on them is often like me just sitting down with the engineer and be like, hey, here's what I think. These are the things I should change. I think part of, there is still time where I feel like, you know, being in the, in the design tools and iterating on it and like getting down to polish and to iteration is like still really, really important.

Mm-hmm.

So that part hasn't like gone away. It just, I think, I think since I'm like operating on more projects at any given time, It feels more effective to sort of just like do it really casually and informally.

Okay. I mean, that's always been like the most fun part of just like, you know, being a PM or designer to me, just like getting with the designer and engineers and just like looking at product and iterating.

Yeah.

So there isn't a lot of, like, do you guys still make like specs and Figma and like, you know, these planning documents or is it just like in the code iterating on prototypes?

I mean, I still make Figma. I don't, I think we haven't, we don't make specs as often. Like they're, and they're not as detailed usually. Yeah. Like we still do that prioritization and it like exists as a document. It's actually really good for us to like hand off to security or legal and stuff like that. So they understand what's happening in the launch, but it's usually just like a few bullet points. It's not this like beautiful table that's over-engineered. And I think the same thing goes for our Figma files.

Okay. So why don't we get into it? So like now Anthropic, and by the way, like I spent my whole life on Claude, like I live in Claude these days, but like now you guys have like 3 separate products, right? You have Claude, Cowork, and Claude Code. And maybe you can kind of like show us like how you use some of these products or like how, what do you use each one for?

Yeah. I will show you Cowork because my secret is that I actually use Cowork for most things now aside from like the, the very, uh, nitty-gritty production code stuff. So like when I'm polishing, I will still use like Cloud Code. Um, but I, for, for all, for all purposes of chat, I actually fully use Cowork now.

Cool.

Um, so I will show you some things there. Okay, I'm gonna show you this account that I'm using with Cowork. There's no, there's no tasks there because this is sort of like not my burner, but it's my, uh, Yeah. It's my external account. Uh, but trust me, like, my actual Cowork has like many different sessions that I'm often like running in parallel and whatnot. Um, and so one thing that I find Cowork really helpful for is I, I just generally call it sort of like garbage in, treasure out. It's just really good at taking things from a lot of different sources and then like finding and like picking the gems out of it and also making it something really productive. Um, so right now I've actually like hooked it up with a folder of, uh, they're, they're not real user interviews. I like actually cut and clawed to generate them, but they're, let's just say they're like a bunch of like transcripts from user interviews that we've done from UXR.

Okay.

Um, and I think the thing that we try to do, at least on the cohort team, is we try to stay super close to the ground, both through like traditional sources of like UXR and like talking to people, um, but also through like internal dogfooding, through, through things like just this running Slack channel of all this feedback, but then also, you know, social and, and people who are using it and passionate users there. And so, um, like, I think that's actually been a big part of our, like, somewhat success so far is that we have just been, like, staying super close to that and iterating really quickly. So what I will do here is, you know, I will say, like, cool, like, I have this, I have this folder of, you interviews, but I also will tell Claude like, hey, can you look on like social and all, you know, like customer reviews and whatnot of Cowork and tell me what the biggest insights are? So look in this folder of UXR interviews and on social like X and Reddit. And other reviews for Cowork. Or I guess I'll say like Claude Cowork. Tell me what the main insights are. And I'll have Claude do that. It might take a little bit of time because it's like going, it's truly just going through so much data.

Yeah.

And processing it. But yeah, it'll do things like it'll sometimes like spin off subagents to do that in parallel. And It'll search the web for a while and do that. Yeah. But yeah, it's basically just pulling from a wide variety of sources.

This episode is brought to you by Replit. Replit's new Agent 4 is the best way for teams to collaborate with AI agents. There's a new infinite canvas where you generate design directions, compare them, and take your favorite straight into a working app. From there, you can spin up multiple AI agents in parallel to build different features at the same time and track them on a task board where you can review and approve what ships. And once your app is built, the agent can generate slide decks, animations, and more from the same project because slides, animations, and documents are all just code under the hood. Check out Replit via the link in the video description and use my code PYANG26 to get 3 months free. That's P-Y-A-N-G-26. Now back to our episode. And in your real job, do you have some sort of a like weekly insight report or something that, you know, gets sent to you and your team or something? Aggregates all this stuff automatically?

I mean, we can make that right now through CoWork. So I'll show you the flow for that. Yeah, we do in that I think one of our researchers has one that they'll send out, and then we have one that sort of like pings us in Slack as well that will send that out. And then we just, we also just like listen to the Slack, like the internal Slack. Like we have really been relying on our internal and like most powerful users to sort of like be the people that gives us like the sort of bleeding edge feedback because people internally are really willing to be honest with you.

Yeah.

And they're often pushing the capabilities furthest. And it's easiest to follow up with them too because they are just, you know, so close to you and like, and willing to participate. So we rely on internal feedback a lot, which might be counterintuitive.

I mean, that, that, that's what should happen. Like, and I feel like most companies, like the, it's like so siloed that like the teams don't dogfood each other's product, but like, it feels like Anthropic, that very much happens, right? Both for Cork and Cockle.

Yeah. Yeah. Like it was a big part, I think, of Cloud Code's success too, is just like listening to your frontline users. And I, and that was the thing too, that we did a lot at Figma too, was a lot of internal dogfooding. Yeah, because a lot of people will point out things that, especially when it comes to things like interaction design and polish and whatnot, people internally will really poke at that stuff versus people externally tend to give you feedback often about like the, you know, does it work for their user flow? So it could just be like a very different altitude of feedback.

Let me ask you this while this thing is generating.

Yeah.

So So I know that Anthropic, like marketing, PM, everybody was like, is basically using Claude Code to also do stuff. Now that Cowork is internally available, how do you see different types of functions or like, how are people using this versus Claude Code or depends?

Yeah. I think we're seeing just like overall, like wider adoption of Cowork for similar use cases that some of like the bleeding, what I call like the bleeding edge users of Cloud Code were. Like, for example, when we started working on Cowork, we had a few internal sales employees just be sort of our inputs into what we should build. And there were a few people who were like, oh, they're like diehard Cloud Code users and they were using it to generate like leads lists and help them like come up with Scripts and whatnot for calls. And it was like, it sort of blew my mind when I saw it because I, at that point, didn't even comprehend that you could use Claude code for those things. And those folks, they are, they're basically all on Cowork now. And more of their sort of like their peers are actually using Cowork as well. Whereas before, those people were like the bleeding edge users.

Just because like, it actually is nice to see a UI.

And I think so. Yeah, I think that's That's honestly what it takes. I think also part of it is it's so close to the other work too that they're doing. If they're already using chat and they can still use Cloud Code too from this desktop app too. I think it fits closer to their existing workflows than opening up the command line does.

Got it. Looks like it generated something.

It's generating stuff. It sounds like there is some There's some insights there. Okay, cool. Like, I can take these insights, um, and there's like 7 different themes here. Um, they're different every week, which is great. But I can, I can basically tell it and say, um, you know, like, it's created this, it's created this docx for me, um, which actually is like, it will be, it's already like stored in a folder, um, on my computer, which is great. Um, and so what I can also do is like, I can spin off 2 parallel tasks. So here Here I can say, okay, cool, these are great insights, but what are product features I should actually build from this?

Yeah.

Then what I might also do in parallel is, it's already attached to this folder that this output has been dropped into. I could say, okay, given the insights, in this folder from the insights doc you made me. Turn this into a presentation I can share with the team this week at our kickoff.

Mm-hmm.

So in one thread, it's making me an artifact. The other one, it's starting to actually figure out what to do next. And I think this is gonna run for a bit too. But ultimately what I can do from here is like start the design process essentially. Like it will give me options for what the features I can create. And from there I can even ask Claude to like create some wireframes for me for those given features. And that can be a really good place where it, it just shows me a bunch of different options and I can take those Bring those with me to Figma and actually start to flesh them out or bring them with me to Claude Code and start to make them real with our real design system components and just get started from there. And then the other thing I can also do too is just take these and turn them both into scheduled tasks.

Yeah.

Exactly.

Yeah.

So I can say, Actually, I probably wouldn't start from here. I would, once this is done running, I would say, okay, like schedule this for me. I don't wanna disrupt while it's doing this thing.

Yeah.

But I would probably just like have it structure, schedule this task for me to do it every Monday morning at like 10:00 AM or something. And then that way I start every Monday morning at 10:00 AM with this presentation and with like, you know, 3 different product ideas that I can use. And can kick off the week 4. And it just makes that sort of like iteration cycle between feedback to actually creating something that's tangible or an idea that the team can look at. Just really, it squishes that time to be really tight and it helps us just iterate on the product quickly and make it much better quickly.

Yeah. It's all about the iteration. It's all about the iteration.

Yeah.

Yeah. Yeah. And you know what? Like I've also gotten so lazy. Like I always let AI take a first Yeah. So if you actually want me to take those insights and try to draft like some sort of a feature prioritization from scratch, it'll take me like way longer than it used to take.

Yeah. Yeah. I mean, that's how I operate too. Like even when you sent me over like the notes for this podcast, I was just like, I, you know, I keep a folder of like all of my personal notes, whether it's like things from one-on-ones or things from, um, just like random thoughts. And then I was just like, cool, like, you know, read my personal notes and like, like come up with, with speaking points for, for this podcast, um, and, and help me think about what I want to say here. And it— I mean, I, I'm not gonna— I'm not reading it word for word or anything, but it just helped me jog my mind in terms of like, oh cool, like these are the things that I, I care about and, and, and I didn't have to think that hard to actually help evolve my thinking instead of just being stuck there with the blank page problem.

Yeah, yeah, yeah.

Yeah. Okay. So, it's doing the thing where it's speccing and telling me what the P0s and P1s are. So, let's say I want to build out the step-by-step task progress UI. I can just say, I like the step-by-step progress. The step-by-step progress UI idea. Um, make me an interactive prototype of that, um, of a few options of this and do it in a scratchy wireframe style.

Scratchy wireframe.

Yeah. Yeah.

Got it. Okay. Okay. Oh, so, so it has context of what cowork looks like now.

This probably doesn't because I haven't set it up to have that.

Okay.

But it will just, I think based on the context here and whatnot, it will try its best. And it feels pretty isolated too. So I don't think I have to show it like, okay, this is what co-work looks like. But I could upload an image and also give it this Anthropic branded skill that we have to make it look more on theme.

Do you have any, uh, yeah, like that's my next question. Like what kind of skills do you use or do you have any personal skills, uh, for like making these docs and slides?

Yeah. Uh, we have a few internal skills that we use for some of these like docs and slides because, um, it helps brand them. I actually don't have like a personal skill, like library that I use. Um, I've been mostly borrowing from the ones that we have internally and using them for different purposes.

Okay. Okay.

Yeah.

Because yeah, so like, for example, I have a skill, like a writing skill that just tries to tell it not to like make AI slop words.

Like, yeah, yeah, yeah.

You know, so that helps a little bit. But yeah.

Yeah. I've actually found that like now with, this is probably not like the best practice and the most efficient thing to do, but with, um, with Cowork and some of these like folders I have with all these like personal notes and whatnot. It's, uh, sort of like learning about me from those folders has been really useful to a way, in a way to me where it's like, it, I, I feel the need less for, for things like memory and for skills because it sort of has this knowledge base about me already. Um, even though I think, I think there's still a, there's still a time and place for skills, but I personally, for, for the use cases that I've had, have felt like less of a need for them.

Oh, because it, it, it just updates as a memory every day. Based on what your conversations. Yeah.

Or it's, it's, yeah. And it's basically like, that is like a, a memory that almost that I, I maintain of it because I'm taking notes in it all the time. Yeah.

Got it.

Yeah.

Yeah, this is great. It looks like your presentation is ready. There's like a dot.

My presentation is ready. Let's see if it is. Oh, it is. Um, okay, cool. So it's made this presentation. It will load a little bit. Um, yeah, so I think, you know, with this you could take it into, into Keynote or whatnot. It did. I didn't ask it to use the Anthropic skill here, so it's sort of generic, but I could. And then yeah, I think from here I would take this and I would say like, look, schedule. Yeah, do this every morning or every Monday morning at 10:00 AM.

And you can set it up in a way where, can you set up in a way where it actually shares with your team too? Not just you, like in Slack or something?

Yes. Like I could, I could set it up with my like Slack MCP and have it just Like send it out, um, basically. So it's like, I think I would, what month? Yeah. So I think it, I also like how it asks me like clarifying questions here. So, uh, yeah, it will, what should this money, let's say it rebuilds the deck. Yeah. So it'll, it'll do that. Um, yeah. And then I could also ask it to, um, you know, using the Slack MCP, send it to a certain channel every, every week.

So, so when do you bring your team into this process or like, you know, mm-hmm. Like, yeah. When, when, when do you actually, cuz you're interviewing with AI. And then you go interview with your team and you kind of go back and forth, or how do you—

Yeah, I mean, here, like the, the, the actual like UXR interviews are like something that I probably wouldn't, you know, like either the PM or the researcher on the team or somebody else on the team would actually do. Yeah. And then we would, and then through this, you'd probably just like share the artifact and bring them in. And then, and then this actually could become the the thing that the team operates off of. It's like, oh, the team, at least our team is pretty bottoms up and democratic in that way. And so giving, the way we operate is we just give people the insights and the goals and then everybody sort of goes off and makes prototypes and tries stuff out and ideas come from everywhere. And so—

I see.

That's what the handoff point is. It's less of like, Oh, like me as the designer, I will come up with all the ideas, but it's like, hey, here are, you know, anybody can do this. It doesn't have just have to be the designer either. It's like, here are the insights. Here's what we're trying to do this month. And how can we all get there? Like, how can we all meet this goal together?

Yeah. Maybe just give like half the tasks to Claude to create a PR for. And then for the more complicated ones, you can like try to handhold it.

Yeah. I mean, yeah, I think with this, we've, We still wouldn't hand it like directly to Claude to do all this stuff. Like a lot of what we do is I think we rely on ourselves for a lot of our own judgment and our ability to like curate and decide what to actually build and do.

And like, you know, people could talk about taste and stuff and judgment online. Like, I don't know about you, but like, I feel like the way to build this stuff is just to, yeah, like just get the firehose of product feedback from internally and externally. And then you just kind of develop this sense of like what's broken and what needs to be fixed because you're just hearing feedback all day long.

Yeah, yeah, yeah, yeah. So this one, I think, yeah, like, you know, it will have created these like basically kind of wireframe-y type things.

Got it.

Yeah, and I think it's like it'll show you different options and stuff like that. And I think I am, as a designer, I like to just like see a lot of options even if they're like not super high fidelity. Versus imagine them, and that helps me decide what to actually do. So yeah, I think having Claude just come up with the options saves me a step from just having to mock them up, even though from here what I would probably do is choose one of these directions and then start to microiterate on it, or even take this and make a prototype of it in code and then iterate from there.

Got it.

Okay.

Yeah. Wow. It sounds like a fun job.

Yeah. It's a fun job.

It sounds fun. Let's switch topics a little bit and let's talk about how Cowork came to be, you know? So, so like there's a bunch of stuff out there about how it came together in 10 days.

Yeah.

But I think it's actually, there was a lot of iteration before that too, right?

Or like, yeah. That's a, that is one of those things that I think got pulled out of a quote somewhere on somebody else's like interview or press tour. And I think everyone just anchored onto it like, oh, they made it in 10 days. And the actual, like, the actual story is that cowork or this, like, direction of cowork has been something that the company has been, like, thinking about or, like, wanting to do for honestly, like, basically ever since I joined Anthropic about a year ago. Just some way to help and be sort of like a thinking partner for— since we already have this basically for code, but for all general knowledge workers.

Yeah.

And I think a lot of it has just been like, how do we execute on it? And what is the right architecture for it? And what is the right UX for it? And so Even throughout last year, there's been lots of different prototypes, some more ambitious than others. There's been a lot of different technical experimentations around different agent harnesses that we've used, a few that didn't pan out super well before we landed here. And we've seen prototypes from people in our labs wing, but also prototypes that we've built on the product side of the house. And I think about a lot of these things, it's just about if the idea keeps coming back, And there's like energy each time, but then it's like the execution. Often it is just about like timing and execution, just like the lightning striking really well. So when we did decide to ship this, it was basically 10 days from like, oh yeah, we should ship this to like, okay, it's shipped kind of thing. And a lot of that was like building on the momentum of what we were seeing with Claude Code over the holiday break, where it feels like over the holiday break, you know, everybody finally had the time. to try out Claude Code because they've been hearing about it. They're like, okay, it sounds great. You know, there's all these AI tools. I don't have any time to try it. But then for some reason, everybody tried it out. And then a lot of also like people with non-technical use cases tried it out. Like people were using it to like do things like parse their podcast transcripts, or they were using it to do all these like complex analyses and stuff like that. And we were starting to see like even like early product market fit of like the Claude code, basically agent harness with non-technical people. And so we already had this sort of working prototype of that internally, and we were slated to launch it a bit later, but I think that gave us the— we were like, this is the moment, we need to meet it. And even if we don't have the perfect product yet, we should get something out there because we think that the benefits and the usefulness and our moment to Capture this audience, it is right now. And so that's, that, that, that ensued a really hectic 10 days, but it got us here. And now that it's out here, it's been really, really helpful for us to get the signal from people of how to use, of how they use it and iterate on it and keep making it better for people.

Yeah, you can learn way faster. Like you just, I guess the research preview is like the new beta.

Yeah.

You can learn way faster. Yeah. Okay. So basically, so if I understand, there was even over the past year, there's a bunch of like people sharing prototypes in Slack internally and then, you know, people leaving feedback and then eventually you had a working prototype and then because the market kind of demanded it, you kind of just like rushed, you kind of just like did a sprint to make it happen.

Yeah, basically. And like, we were originally going to launch this a few weeks later anyway, but it felt like, oh, this is the moment. And it sort of forced us to scope it in a way that was a little bit more realistic. And like, and, and just like put energy behind it and staffing behind it.

And can you, can you share some early iterations or any kind of like work in progress stuff?

Yeah. Yeah, I pulled together a few screens of what it actually looks like or some like design iterations too of it. There's obviously like a, there's a lot more, but this is what I can show you. Okay. So we had like an earlier prototype in the year. And this is actually something that another designer and I collaborated on. But we actually tried to make this sort of feel a little bit more like task-oriented or like workflow-oriented. And I think also we were really worried at that time, like, oh, like Claude is— or people are not going to understand that with a product like Cowork that they're going to be able to you know, do certain things or like make certain basically outcomes, like make a dashboard or something and gather sources from a bunch of different places. And so we actually made this the UI much more structured and almost like you would like a workflow tool where it's like, hey, add these things. Like, these are the inputs and these are the outputs.

Right.

And then we made chat sort of secondary here. And I think when we did this, like one, Claude was like not great at sort of following the workflow exactly at that point, given the technology that we had. And it felt like way too structured. Like I think seeing this UI on the left where you sort of have to fill it out and like add all these different sources, it didn't feel good, you know?

Yeah.

And it was just like, it felt like a lot of work and it was like, why do we have to do this in this day and age in 2025? When we could just ask Claude to do something for us.

Got it.

And so that was one iteration that was actually a part of a different prototype.

Yeah.

Got it. And then we originally were just like, okay, we eventually were like, okay, cool, it's going to be a chat box. And we tried to do a thing where we guided people towards more specific outcomes, whether it was analysis or documents. And we built this working prototype that I wish I had a screenshot of where each of these, when you clicked into them, they almost had different dials for you to adjust the length of the document or—

I see.

Or the type of document, like memo, presentation, whatever. And that felt really overwhelming as well, I think. And so a big thing we were trying to balance with a lot of these explorations was how prescriptive of the use cases are we versus freeform, like a chat box.

Got it.

And then what we actually ended up shipping with just a few weeks ago, a few months ago, was something like this. Where again, we did have this sort of almost wizard-like experience. We would click through and it'd be like, oh, create a doc, make it 3 to 5 pages, et cetera, et cetera. And we also just showed a bunch of the UI up front because we wanted it to feel like really differentiated from chat. Um, but then I think when you come here, it just felt like there was a lot of competing visual elements. Um, and so over time, like, we, we stripped most of this back. Um, and, and, and we went away from this actually like really opinionated UI because, uh, it was actually like not helpful to show all of these things. Um, and so yeah, it's this constant balance of like, how much do we tell people how to use, what to use it for versus like leave it really freeform for them. And so now actually, like if you look at the current UI, which I'm just gonna switch. If you look at the current UI, yeah, like a lot of it is really peeled back. We don't have, we don't, we're not showing you the heavy sidebars and it feels more like the traditional like sort of chat box. But we, we sort of have, we reformatted the homepage to feel more almost like It's Claude's active to-do list. So here I only have like one active task, but at any given time I could have like many more. You know, like if I'm really asking Claude to do stuff or if I've like, you know, given Claude a bunch of tasks and come back, I'll have a bunch of these where it'll be like, okay, there's an extra message or I could review this. And here I can actually just like approve and triage these tasks.

Yeah.

So it's like, yeah, let's schedule that. And then I'll see my scheduled tasks here. And the idea is that it, yeah, it just feels much more like a shared to-do list between you and Claude, as opposed to this chat box with like a bunch of overwhelming suggestions and UI to sort of show you how to use it.

Got it.

Yeah.

That's awesome. That's awesome. And maybe in the future it can be like, you know, like you have like multiple agents and they can have like a Trello board. You can like move tasks around and stuff like that. Yeah.

I mean, maybe, maybe, yeah.

Maybe by the time this podcast comes out, you have updated the UI again.

Yeah. Yeah. Maybe, maybe. Yeah. But I think that also shows, hopefully that shows too, like the UI looks so different, just like, you know, maybe like 4 or 5 weeks ago. And, and now it's like, we're, we're constantly learning like what's working and isn't working, but we're also learning best about how to like display This technology to people. Yeah.

Yeah. You know, it's really interesting, like, for ClockCode, like, I've been leaving, I've been like tweeting some feedback about like, because there's so many different like slash commands and stuff that you can just like learn. It's almost like, it's almost like going to Costco to do like a treasure hunt. Like, you never know what you find. But like, but like, it's not very beginner friendly, right? It's just like, it's like a game where you like learn more and more and kind of master it. I think maybe for Cowork, you're trying to find the middle ground between, between the regular chat and the cloud code. So it's not like all the things are hidden, but like, you know, you try to guide people in some way.

Yeah. I mean, we, you're still able to use slash commands in Cowork, but they're not like the primary way in which you interact. And I think the way that at least I personally think about it is, you know, at least Cowork is also a tool for professionals and We sort of see a bunch of people use it in very power user-y ways, and there are already power users of it. So a lot of people will be motivated to learn much more complex capabilities, right? Like, they will be willing to create their own skills, share them with their team, etc., and also want shorthand, you know, commands for things. But most all these things should be possible without having to do that learning as well. And those sort of commands should be a secondary way to interact with Claude, and you shouldn't have to know all the commands in order to use it.

Yeah, that makes sense.

Yeah. Okay.

All right. Well, just a couple of closing questions, I guess. So how does planning work at Anthropic? It seems like a pretty chaotic place. Do you guys do annual planning and setting goals and stuff? Or Is it more just like prototyping and trying it out?

I mean, it also changes every time we do the planning. Like, I feel like every time we've done it, it's been very different. On the team that I am on, we do monthly planning. We just, we have a spreadsheet and I think it, at least in the cowork section, there's maybe like max like 12 things in that list. And it's really truly our P0s. And then for each of them, they just have one single DRI and it's like, And then we just come back each week and we're like, are we on track or not?

Got it.

There is some amount of like quarterly or half planning that we'll do where I think usually one of the leads will be like, hey, this is like generally where I think we should go and these are the things that we should work on. But it's not so structured where it's like we absolutely have to do these things and these are the projects. It's sort of like just giving everyone a picture of how things might fit Yeah, yeah. So it's pretty loose.

It's pretty loose, right? Yeah. You know, that's the funny thing I find, because like, you know, Anthropic is one of the most innovative companies. But I find at the most innovative companies, it's less about trying to do this annual planning theater and more just about like iterating and like learning from the users. Like, do you still make like, at some point in your career, you probably made like a North Star vision deck or something, right? Like, do you think that stuff is useful?

Yeah, yeah, yeah. I mean, I think we did— I made a North Star vision deck maybe last year.

Yeah.

So I think there is still room for these visions in that they point people towards something and create clarity around what we're going to do. I think given though where at least I work, this technology literally changes all the time and there's new models all the time, and the rate at which new models is coming out is just faster and faster, basically.

Yeah.

I think there's no such thing as a 1-year vision, let alone a 2 or 5-year vision for us, because there's so much that's unknown out there. But what actually is helpful still is pointing people towards the same direction, especially when anyone can build anything. So what I think about as visions now is just more like, Yeah, it's like 3, maybe maximum 6 months. But a vision can be— it can be a document. I think it's more helpful when it is visual. So I do think this is like where design still has like a ton of power is being able to piece things together and make it coherent in a story for that period of time. But it can also be a prototype. Like it doesn't need to be like, yeah, this like deck that's static. It can be a prototype that just sort of helps us I think often what it helps us do is like we often have, you know, 5 teams working on something like very similar or things that collide or might feel duplicative. And what design can really do is like curate and help bring cohesion to those ideas and show us a path towards like making that an ideal experience as opposed to 5 disparate ones.

Yeah, that makes sense. That makes sense. And how about like with, do you guys have like reviews of Ami and like for folks like that? Like, is there like formal reviews or? They're also playing with the prototypes too.

No, we do still have reviews. We don't have them for— I've been at companies where we have them for every single feature and whatnot. It happens for sort of bigger, higher priority projects. But it doesn't feel like it is a big sort of thing you have to prepare for and spend so much time on. It's mostly for that visibility and for that feedback. And if there's things that sort of cross our big and have like big company-wide implications, we'll have those reviews. Yeah.

Got it. Okay. Well, I mean, do you have any closing words of advice for designers who maybe like, you know, they kind of feel like the ground is shifting under their feet? Like, should they start like submitting PRs or like what should designers do?

Yeah. I mean, if you feel like the ground is shifting beneath your feet, it's because it is. You know, it's because it is. And I think at this point, you have to just accept that, you know, and then sort of adapt to it and just be really open towards like questioning the way that we work already. I think it's coming for designers right now. And I think we're feeling it today a lot because a lot of like the roles around us have changed already. And we're like, the second-tier effect for it. But also, I think the tooling for us is also changing.

Yeah.

And whenever I feel sort of like, sometimes I am like sort of threatened. I'm like, oh my gosh, my job is changing a lot. Like, people don't value me in the same ways that I used to be valued. But then I think about like my engineering peers and how much that they have already adapted towards how much their job has changed and how like valiantly they've gotten through it and how much they are still They're producing even better and more work today. And so I look at them as inspiration. I'm like, oh, like if all these people that I really respect can do it and do it in a way with such humility, then I can too. And I sort of see them as a model for like what's coming next for me.

Yeah. In some ways you get rid of, there's a lot of like mundane and busy work, like moving boxes around that you don't have to do anymore. Right. Like hopefully, so you can actually focus on a higher level thinking and, you know, that kind of work.

Yeah. Or just producing more work too. You know, like I think I, I think about how much my engineers are capable of right now and I'm like, yeah, that's, that's wild. Like they're, they're, they're creating entire features in days, not weeks these days. So yeah.

Yeah. But the funny thing is you don't actually get more free time. You actually work harder.

I mean, like it's because I think we're all really ambitious and we like—

There's like so much to get done.

It's like a high. You're like, oh my gosh, I can get so much done now. And I, and I don't have to do the stuff that I didn't like doing before. And so let me do more, you know?

Yeah. Yeah. Yeah.

Yeah.

No, that's good. That makes more job more fun. Okay. Well, thanks so much, Jenny. This is awesome. Awesome. I feel like we've been talking online for a lot and I finally got to talk to you. Yeah. It's been great.

It's been great to have this conversation.