Loading summary
Transcript174 lines
- [00:00]
Dan Shipper
If you're a non technical person, you are used to a world where you send a prompt and then you get a response within a couple minutes. And once you send a prompt or a chat, you can't do anything else with that AI. This is built for working with your AIs in an async way. This new Claude Cowork app is a really good example of agent native architectures, which means at the bottom of the app, instead of having like software that works by deterministic rules, you have an agent and the agent is wired up to the UI of the app.
- [00:29]
Felix (Anthropic Team Member)
I've been at Anthropic for a little bit, but this is the product that my team has built here. We sprinted at this for the last week and a half.
- [00:37]
Dan Shipper
What we're trying the last week and a half.
- [00:40]
Felix (Anthropic Team Member)
That's it.
- [00:41]
Kieran
Come on.
- [00:43]
Dan Shipper
We've got a new Anthropic drop. So Anthropic just dropped Claude Cowork, which is basically Claude code for non technical people. We got access to it early at every and so I'm going to give you a quick run through of what it is and how it works. We will have a full write up on every. In a. In a few hours probably. We're just, we're kind of figuring out how to do these things. So I'm going to add. Kieran. Kieran is here. Hello, Kieran. How you doing?
- [01:16]
Kieran
Hey, what's up?
- [01:18]
Dan Shipper
So I was just telling everybody if you just got here, we are about to do a vibe check of Anthropic's new Claude Cowork feature, which is basically Claude code but for non technical folks. I'm going to just demo it for you right now in here. Let me just share my screen. This is what it looks like. It's Claude Cowork. You'll see. We've got the chat on over here to the left. You still have got regular chat, you've got code here and then we've got Cowork. So it's, it's three. The three C's it starts with. Let's knock something off your list. I love the copywriting here. You can tell from where they did this that the, it's, it's really designed to do deeper tasks on your computer than maybe chat is. So, you know, create a file or crunch data or make a prototype, send a message, organize files. It's got over here progress. It's got artifacts, context. We've been, we've been playing around with this for a couple hours now. I think it's cool I think it's cool. I think there's. There's a lot of like, really interesting questions from the UX perspective of chat versus code versus cowork. And I think you can, you can really think of coworkers being. It's like chat that has access to your computer and runs for a long time, which is essentially flood code, but just less intimidating. A couple of things that we did that I think are interesting to look at. Here's an example of it working. I asked it to go to the every to website and find five competitive companies that do the kind of consulting that we do and then analyze our positioning and you'll see it just went and used my computer. It's running in a long loop. So this is, this is a really interesting one where you can do this with regular Claude. Like Claude can do this. But the number of iterations that it's going through, this is many, many, many minutes of iterations. So it looks a lot more like. Actually, can you make this markdown? So it looks a lot more like cloud code, but it's friendly enough for anyone to. Anyone to use. We'll look at the. Another thing I had to do, I have a. I have an. I have a dinner that I have to go to tomorrow night that I had to prepare some remarks for and I asked it to basically like go, go to my Gmail and prepare remarks. It has a. It has a connector to Gmail and to Google Calendar. But the connector wasn't working, I think because this is like very beta and was not out when I was testing this. And. Let's see. And I asked it to draft a response and it drafted a response and I think the response is actually pretty good. So this actually sounds like me. This is kind of crazy here. And you got to look at this because this is implications for Quora. So basically the setup was I have a dinner tomorrow that I have to do remarks for. And I asked. And the person who organized the dinner was asking me, can you tell me like what you're gonna. What you want to talk about? And so I just said like, read, find the email and draft a response based on what you think I would say. And this is something that I could, I think I actually could send with minimal edits, which is. It's pretty cool. Are you seeing this, Kieran? Yes.
- [04:42]
Kieran
Looks good.
- [04:43]
Dan Shipper
I think one of the things that's good about this is it's gone through many, many steps to both identify what I would say and how I would say it and has all the context which Is like. Yeah, it's pretty cool. Yeah, go for it.
- [04:57]
Kieran
Yeah. So how is Cowork different than chat? Like, GoWork is more made to. You can share your screen, but like it's more made to like go longer, like really work on something so it's more focused on getting some work done. And this is what we know in close code already, like when you trigger ultrathink or your trigger planning modes, things like that. So it is similar in certain ways, but it also, it will unlock a lot of things that you use as a developer in close code that now suddenly work very well for non developer tasks.
- [05:32]
Dan Shipper
Yeah. I can sort of see this being a thing that just really accelerates our growth team or our people who are in our consulting business who want to do some of these tasks and maybe are using cloud code, but maybe it's a little bit less intuitive. It should be released now. I think that they're maybe holding it back. But the blog post is live, so you can take a look at the blog post. It's Cowork Research Preview. I'll throw it in the. In the chat and. Yeah, so I think it'll be out soon. Let me. I can check with my anthropic people, but. Okay, let me go back to this. So I also asked it to do a calendar audit. This looks like it's actually still running on this, which is crazy. Asked this like an hour ago. Go through the past month of my calendar and do an audit. Tell me how this reflects my priorities and whether it's aligned with my goals. Just browse on Chrome.
- [06:31]
Kieran
Can you share your screen again, please?
- [06:32]
Dan Shipper
Oh, shoot. Yeah, thanks. Yep. Tell me how. Go through the past month, my calendar and do an audit. Tell me how this relates to my goals. And it just, you know, it's been browsing on my computer for like hours and hours. One thing that's. Or not hours and hours, but like for about an hour. One thing that's different about this, which I think is really interesting, is in Claude, when you start a chat and it's responding, you have to stop it in order to send a new message. But this just output. I can just add to the queue. So this is a little bit more like the. This is more like the cloud code experience. So I think this is one of those situations where it's built for you to send stuff as it's working and it's not like one message, one response, one message, one response. So it's really more built for long tasks.
- [07:30]
Kieran
Yeah. And it has the todo task also built in on the right, which is nice. So you can see where it is and what it's doing.
- [07:38]
Dan Shipper
Yeah, exactly. Another thing I did is I fed it a book and I. This is a book that I. It's called the Outsider that I've been reading for a book that I'm writing and I just asked it to like basically read the book, read the entire book and construct a taxonomy of all the main characters and ideas. And looks like it did this. And this is something that you could also do with, with the. With regular Claude, but it would be. It would just be less detailed. This is really interesting. I wonder.
- [08:12]
Kieran
So, yeah, so if you want to trigger those longer running tasks, like you can just say, make a plan, do this. So I assume if you just push it to do that more, it will just run for longer. So if you say take an hour, read through every single email, it won't give up as easily as before and it will just keep going. So please try those things.
- [08:34]
Dan Shipper
Yeah, it does have a plan mode. It said I have not used the plan mode yet, but that's pretty cool. I also asked it to go through our post HOG analytics and do some data gathering. So we published this guide. If you haven't seen this guide, it's like I keep talking about this. Everyone at every is like making fun of me because I can't stop talking about agent native architectures. We published this guide last week and we have these buttons, Read with Claude, read with ChatGPT. And I was curious, okay, how many people click these buttons? And that's the kind of thing where I would go ask Andre who runs our platform and be like, can you go look this up? But instead I just said, can you go into post? I just went to Cowork and I said, can you go into post HOG and just find all this stuff? And it said, okay, Total chat with Claude Button Cooks 4000. That's actually freaking crazy. Oh my God.
- [09:28]
Kieran
We should get money. We should referral.
- [09:31]
Dan Shipper
Referral fees, baby. What about chat with ChatGPT? What about copy for agent?
- [09:40]
Kieran
And so this is cool because we. We were in a rush since we started this morning and we didn't have any MCP set up. What we just did was we connected Chrome and Dan is logged in on Chrome in post hoc. So just browse through the thing. Got the things. So that's very handy, like mvp. Just make sure you connect Chrome. That's a very good one to add already. You can also do that with the normal. But it's very good at browsing and Figuring things out, especially now it doesn't stop as quickly, which is really handy. So anything you can do in a browser, you can now use Cowork for to use longer running tasks and kick off things. And you can use multiple tabs as well. So you can have five tabs being controlled by Claude and running, which is great.
- [10:33]
Dan Shipper
Totally, totally. And it sounds like this is down for people, so come hang out.
- [10:39]
Kieran
For me, it's done for me as well, but. But for Dan it's working.
- [10:44]
Dan Shipper
So I'm the only one it's open for. So we've got a monopoly here on, on anthropic Cowork content. So if you're here and you have stuff you want me to try, just let me know. I'm happy to, happy to throw it in the chat and, and make sure I have. I have something fun to do. Make sure that you use every you read every because every is the only subscription that you need to stay at the edge of AI Every to. We've got these vibe checks. When new models come out, we get them beforehand. We had this several hours ago. We knew it was coming since last week. And we always have all the up to the minute stuff that you might need. We also have a bundle of apps that we make. We have an app like the one that Kieran makes called Cora, which is an assistant and email assistant. We've got one called Sparkle that helps you organize your files. We've got Spiral, which helps you write. And we've also got Monologue which is a speech to text app which I will show you shortly. And let's actually go back to. Let's actually go back to the Claude demo real quick. So. And Karen, if you see anyone asking questions just.
- [11:53]
Kieran
Yeah, yeah. There's one good question from Hunter. He says how good is it at research? Like one of the things I love Cloth normal for is the deep research. Like does that exist? So maybe we can see if it can do deep research on something.
- [12:09]
Dan Shipper
Yeah, Let me know, let me know if you have a research query you want me to try. But like I can show you. You know, actually, you know what, I'm going to share a different screen. Hold on. Plus share screen. Okay. I'm figuring out my live stream setup. This is the first time that we've really done a live stream for a vibe check. Okay. So now you should be able to see my screen again. So for research, like. Okay, it depends on what you mean by research. Right? Because this is a research query. Can you connect to Posthog and tell me for the agent native guide that we post last week, how many people clicked the chat? Like that is research and it does give me actually like a really good answer, which this is so interesting. Okay. But like another form of research that we talked about is okay, analyze my competitors. This is specifically analyze our competitors for the every consulting business. And I'm going to say open and proof. Proof is the agent native markdown editor that I built over the weekend. Can you believe I just said that? Um, and so this is a, this is a research document that Claude Cowork put together. Which, you know, it's not like it's not the meatiest thing I've ever seen. Let's see. I think this is not bad. I'm not noticing like a super significant difference between this and like what a normal clod would pull. Pull out. But I can see the research itself is more, is much more extensive than normal Claude would do. And let me just actually throw this into normal.
- [13:56]
Kieran
You can see also. So if you do deep research, the research agents in chat, that's pretty extensive normally, but here you can see more. So yeah, I don't know, maybe that is also available in that version. We're still figuring out what everything is that is available, but it's very close to what you can do in Claude code. So there's a kind of a fusion of chat and Claude code is cowork.
- [14:26]
Dan Shipper
Yeah, yeah. This is very like. You can see that they're calling these tasks as opposed to chats. So it's supposed to be. I think here's a good way to think about it. A good way to think about it is if you're a non technical person, you are used to a world where you send a prompt and then you get a response within a couple of minutes. And once you send a prompt or a chat, you can't do anything else with that AI, you have to move on to something else. This is built for working with your AIs in an async way. So everything is set up like the idea of a task, the idea of having a queue. This is all set up so that you can, you can set, say go do something and then not think about it for a while and then come back. Which is very different from Claude where the normal cloud app, you're, you're kind of, you're, you're trying to, you're trying to get an answer pretty quick. And I think that's the best mental shift. I think the real question that I have is, is this deserving of its own tab for One. One reason it might be is there's a difference between, like, how you might treat one of these versus one of these. Like, these are more throwaway. These are probably like bigger chunks of work. But honestly, it's like kind of confusing. I would rather just. I think that they. I would rather just have it all in one tab and then have it do different levels of research and thinking and. And async based on the task and maybe based on, you know, a setting. Like saying, like, really fucking think about this. I don't know. What do you think, Jiren? How would you solve this?
- [15:58]
Kieran
Yeah, yeah, like, for me, it's also confusing. It's like, oh, great, there's another tab and I have to like, first think where to go. But I do get it because I've seen this transition as engineers, as an engineer, like, we had the, like, copy paste into ChatGPT, obviously, and then like, that evolved into cursor, like more agentic, which evolved to like, I don't look at code anymore. And I think there will be a similar transition for people that do research or cowork. Like, maybe now people are used to go into Chrome and like, seeing what's going on, what's happening, towards more of, like, I'm just going to let it rip and do its thing and then after it gets back, I'm going to review whatever the output is rather than understanding every single step all the way. And. And then you're like, clearly already there. That's how you're thinking. But I do understand if you're not there, if you're more like in the chat thing, where it's like, oh, do this now. Why don't you look at this? Like, if you have this conversation, it's maybe more chat and cowork is more. You hand off a task to your agent and your agent comes back and you review the work and you can follow up, but you can give extra directions. But I do understand to introduce a new tab because there's. You need to shift your mind to do that and obviously we're doing that. But I understand lots of people in the world, especially that are not coders, still have to make that transition where you just hand off something and then let it do stuff for 30 minutes or an hour and then come back and review that. So I do understand why they want to separate it even though it's all the same technology. Because chat code and coworkers, it's all the same model and it's very similar. Harnessing around is philosophically or how you Use it may be a little bit different. So I guess that's why they did that.
- [17:57]
Dan Shipper
Yeah, it's interesting too because when we did this original vibe check, when we got it a couple hours ago, we had me and Kieran and a couple other people on the phone from internally at every and we were demoing it together and their initial reaction was like, I don't know how this is different and is this like even that, that useful compared to regular Claude or cloud code? Because a lot of them are just using cloud code directly. And I think that that's like. That was a really interesting thing for me where you're not going to actually realize how useful this is until you get your hands on it. And there's probably going to be a learning curve on, on it where if you're a non technical user who is used to, who's not used to the idea that you can just hand off your work and then come back, it's probably going to take a while to actually figure that out and get used to this as a UX paradigm. Maybe there's some benefit then to having it be a separate tab so people can basically realize, oh yeah, this is different and I should treat this differently. It's a real adjustment.
- [18:57]
Kieran
Yeah. If you want to learn about that adjustment, we're writing about that for coding, but you could really apply whatever is happening to coding probably for Cowork as well, like how that shift happens, how that goes. So yeah, we thought about, we wrote about some of these things and I created a plugin around this idea. So what I will do in the coming days as well is see if the pattern or the paradigm of compound engineering, if that applies to Cowork. Can I get that working in Cowork? Because I would be very curious to, to expand that and see, see if it works inside here as well.
- [19:44]
Dan Shipper
Anthony, I see that you're from Anthropic. Do you want to come on the stream? I'm going to send you a link. I would love to hear what you have to say about this and anything that we should try or anything that's missing here. Give me a sec. Copy. Let's see. All right, I sent you a stream link. If you want to come on, feel free. If you, if you're, you know, if you're not feeling it, that's also totally fine.
- [20:09]
Kieran
And yeah, he says it looks or is closer to Claude codes than chats and it feels like that because all the tools like the Ask User Question tool, stuff like that, have a ui, which is nice. So you can ask it to say, hey, can you interview me, ask me a few questions? And there's like a nice UI with multiple choice and stuff like that. So yeah, it's really cool, that ui.
- [20:32]
Dan Shipper
It is really cool. Okay, so let's keep looking at this. It's still working. One of the things I noticed is when it was erroring, it gave an error from Claude code. Like the internal error message is cloud code. So it seems like it really is just like a UI wrapper on Claude code rather than a different agent harness or maybe like a cloud SDK. Anthony, if. If we're wrong about that or anybody else from Anthropic who's. Who's listening, I'd be very curious for your, for your thumbs up or thumbs down on that. But that's like interesting design choice not to use just like actual cloud code. I assume that's because it's already in the app. So it's like pretty easy. You don't need to use the SDK, but it's a really interesting thing to see. One thing to. One thing that I want to show people in case you have not been watching everything that we're doing at every. And if you haven't been, I don't know why you would not. I don't know what's wrong with you. But one thing that's really cool that we're thinking a lot about is agent native architectures. An agent native architecture. This app is a really good example. This new Claude cowork app is a really good example of agent native architectures where we think about agent native architectures as sort of like Claude code and a trench coat, which means at the bottom of the app, instead of having software, you have software that works by deterministic rules. You have an agent and the agent is wired up to the UI of the app. And so when you click a button, it is actually just going to the agent with a prompt. I think this is a new way of building applications that we've been working on internally at every. If you're interested in that, I highly recommend that you look at this guide. We'll put a link in the chat, but this guide goes through how to use or how to build agent native architectures. It's like, it's pretty cool and it's like, it makes you build stuff. Like, I built this. This is a markdown editor that I built with. This is a markdown editor that I built with cloud code over the weekend. So in the last couple days I built this whole thing. It helps me Track. We use it internally. It's called Proof. It helps me track. When I get a plan from Claude, it helps me track, okay, what things have I approved? So you can see I'm approving things here. I can track what have I approved, what have I looked at, what's done, what's not done. There's a lot of cool stuff here. I'm going to stop blabbing about Agent Native and go back to Claude. Kieran, anything you want to add here?
- [23:04]
Kieran
Really what I'm curious for is as a power user, can I load my own plugins and stuff like that? There were questions about can you do custom MCP? I think all the FCPs you can do. And also this app has access to your machine so can use AppleScript to load things on your machine, which is really cool. Yeah. And yeah, I'm really curious for where does it go? Clearly it's trying to get cold codes to the normal user. But I think as a power user I would use this as well because in the morning I start up cloth code to make a daily planning like, well, what am I going to work on for the day? But it feels it's probably nicer to do in an app. And also if this translates to mobile, this is super powerful because you can do more powerful work on your phone that's not necessarily code related. So yeah, we need to experiment but there are lots of interesting things. I'm very curious as a power user, even though it is the same technology, maybe it's a new way to use it, which is interesting.
- [24:11]
Dan Shipper
Yeah. So this is my calendar audit that I asked it to do. So I asked it to audit my calendar and compare it to my goals. I reviewed your entire month of calendar activity. So this is something that like regular cloud probably would not do. I have a lot of meetings and standups, so that's an interesting one. I actually don't go to a lot of these, so that's probably not fair. And I also have a lot of one on ones. I have a lot of podcasts scheduled which I'm starting to get rid of. So it actually did a pretty good job. Content, media, health, non negotiable. Many days have 10 to 15 plus scheduled events. That sounds probably right. This is interesting. It said what are your top priorities? I would expect Claude to know this. I wonder if it has access to all my memories yet because Claude definitely knows like what my priorities are. But anyway, this is pretty cool. I like this. See if it did. Oh, we got, we did, we got the taxonomy. It's it feels like it's. It's like slowly lazy loading all the conversation history. So it doesn't. It's like this isn't happening. This is already. This is already done. It just didn't load it. I feel like a lot of the affordances here, they haven't built the statuses yet. So it's easy to see like, you know, in. In the code tab, for example. Like, I could pretty easily see usually what I've merged and what's waiting for me and stuff, but this is like just an. On an unorganized list. I guess by recency or it's not. There's no visual differentiation, which is interesting. Hello, Kate, our editor in chief. Kate is just off camera. Kate. Things are going well. We have. We've got 2300 people here, so looking. Looking at cloud code together. That's so exciting. Yep. Oh, you want. Okay, let me just. Yeah, maybe I should try that. If anyone from Anthropic wants to come on the stream and talk to us, I can see you commenting in the chat. I'll send you a link to streamyard. Just like say, yes, I would like to come on the stream and chat. We're very, very friendly and I think a lot of people would love to hear from you if you're already here. And I will update the app. This is important. I have a beta build right now, so this is something that was not. I didn't. I haven't updated my app and I'm a little afraid to do it on a live stream. Kieran, do you have it in your app? I'm curious.
- [26:34]
Kieran
I'm trying to get it working, but it was down for a while here, but it looks like it's back up again.
- [26:43]
Dan Shipper
Cool. Anybody have other questions or things they want me to try? I'm taking requests, so. So ask any interesting queries. Most interesting query. We'll put in every. When we do our vibe check. Let's see. I wonder if I could use this to code. I'm kind of like, let's see. Oh, this. It did our visual. It did our audit of every. The Every Agent Native guide.
- [27:09]
Kieran
Can you see if it has artifacts as well?
- [27:12]
Dan Shipper
If it's somewhere? Maybe it does.
- [27:14]
Kieran
Okay.
- [27:14]
Dan Shipper
It does have. Well, it didn't do this one in an artifact, but we do have artifacts in another place. Let me just find it.
- [27:24]
Kieran
I do. The fact that the context is clearly spelled out on the right.
- [27:30]
Dan Shipper
Yeah, that is nice. Yeah, context is here. I saw the artifacts tab somewhere. Yeah, it's just like a friendlier ver. It's like a friendlier version of Claude code to me.
- [27:43]
Kieran
Yeah, mine is working as well.
- [27:45]
Dan Shipper
Can you find my every proof, repo and cascade projects? Basically I want you to do a summary of the new feature, the provenance. The new feature provenance tracking that I've been building in every proof and write it up in a nice HTML file artifact that I can send to current. It's to explain to him how the new provenance is going to work. I think this is a combination of something that is kind of. It's dev work related but it's probably not something I would ask Claude code to do. I don't have access to your max file system in this Linux VM environment. Interesting. So that's interesting because it definitely does have access to. One of the things that gets confusing about this I guess is when it's running on your computer versus when it's not. I think it's unintuitive to the average person probably that when you're using it in chat it is all online and when you're using it in here it's actually on your own computer. I'm very curious how they thought about making that clear from a UX perspective. And if anyone is from Anthropics on this stream, why does it think that it can't access my max file system? Maybe I have to add that folder specifically. Yeah, that's what it is. That's so interesting. I just wanted to YOLO give it YOLO access my file system to be honest with you. Look at all these projects. By the way. This is how you know I have a problem. These are all vibe coded projects basically. Let's see every proof. You know what? It's this one. Okay. Always allow. All right. And now I can. The nice thing about Monologue. So Monologue is one of our apps at every is I can just. There's a shortcut for this. I can just click it and then repaste and Cool. But yeah, what I really want, I just wanted to access my whole computer and it has that. That's interesting. It has that file cleaning prompt. I wonder how that works if it can't access.
- [30:07]
Kieran
Yeah, I'm testing it now as well. I'm trying to see if I can use the skills I already have.
- [30:18]
Dan Shipper
Oh, interesting. So I said organize and tied up my downloads and then it seems like it figured out this is so agent native. It figured out how to select. It's as if I selected that folder in the UI because I specifically asked for it. That's cool. I Think that's a really smart affordance. It just like I wish it had gotten activated here too. Like ideally knows that I'm trying to access a folder in Cascade Projects and it is as if I clicked this. Are you sure it's the natural. See if this is actually working. Felix, are you joining? Amazing. Let me just find you on the X app. X, the everything app.
- [31:17]
Kieran
Then if you can share my screen and I can show you.
- [31:21]
Dan Shipper
I will do that.
- [31:22]
Kieran
Let me share while you do that.
- [31:23]
Dan Shipper
Actually, this is helpful. All right, remove. There you go.
- [31:30]
Kieran
So I was trying this out. Help me generate a VST plugin, ask me user questions and I found a little bit. So I found some things. So basically the ask user question I love because it's this UI, it like runs you through and you can hit 1, 2, 3, 4, 5, which is very nice. The weird thing is I didn't answer it and it started automatically skipping this. Maybe it's fixed now, but in the other one it started automatically. Oh yeah, here, skipping. There we go. See? So if your mouse is not on here, it just thinks, oh, this user is not here. So we're going to skip this altogether. Which is very confusing. But also I love it because I'm a dangerously skip permissions person and I understand. But it's weird because if you're here, scrolled all the way up, it will skip to the next. So just a little bit strange here, but the cool part is here I can say three and it will go and continue. So there's like, I like that it keeps going and it's set to like keep going and finish the skip. UI here is a little bit weird, but let's say multi tap, delay and you can see it's working with my skill or skills, skills.
- [32:58]
Dan Shipper
Choose.
- [33:00]
Kieran
Oh yeah, it is mine. Happy Hopper is just the local place where this is happening. Say this. So this is an interface that never existed before, which is cool. Yeah, it's a little bit weird because it's in line here, but in reality you're answering, so I'm sure. Let's see. Sending requests. Yeah. So this skipping part is very confusing. I don't know why it's skipping now, but I do. Like, I would rather say maybe when I start the session, what kind of session it is. Like if it's like yolo, let's go session or where it like pings me and like very clearly in the cowork tab says like, hey, you need to answer a question here. I need your attention because like there's something to be said to both. And now it's kind of somewhere in the middle where it's not super clear. So I rather have it say yell at me and say, yo, I need your input on something. And I don't want to give input on creating or using things. But if it will change the direction of what this will be with the ask user question, probably that is handy to have.
- [34:20]
Dan Shipper
So far this interesting since we have it seems like there are some anthropic people on here. I do have a feeling about ask your user question. I'm. I'm curious what you think, Kieran. I just. There's a limit to how many characters it displays and then it just goes over and hides the rest of my answer. And that just annoys the shit out of me. Do you know what I'm talking about?
- [34:44]
Kieran
Yeah, I. I know.
- [34:45]
Dan Shipper
Yeah.
- [34:45]
Kieran
It needs to be a little bit more flexible also. Like, I want like, why not 20 options? Like why is five the maximum or something? Like sometimes you just have to 20 that you need to like. I get it. But also. Yeah, stop questions. Go build. So that is nice. Just make it now. Okay, so it did not really. Like it's still doing stuff there, but here, like it's. I mean, it's a little bit wonky still. But I love the ask user question flow. It's very useful. Okay, so it does it here and you see the to do right, which is here on the right.
- [35:27]
Dan Shipper
Wait, I was distracted. What plugin are you building here? Can you back me up? And I think about a thousand.
- [35:32]
Kieran
Yeah. Okay. So I have a skill called the Juice skill which knows everything about VST development. And I'm creating a. I ask you help me brainstorm a VST plugin. And we're doing a delay. Oh. Like it's an audio effect. So digital sound processing that you use in your music making. And I'm making a delay now and it's building the delay. And like normally the building of the delay close code is great for. But sometimes you want to brainstorm, you don't want to build. And that's why I like Cowork, because I just want to brainstorm a little bit. And the cool part is it has these skills.
- [36:17]
Dan Shipper
So. Yeah, I want to interrupt you really quick, Kieran, because we have a member of the team from Anthropic here on the stream. Felix. Welcome.
- [36:27]
Felix (Anthropic Team Member)
Hi, friends. How are you?
- [36:28]
Dan Shipper
Hey, how are you? We've never met before. Tell me about you. What do you do at Anthropic? How are you involved in this?
- [36:35]
Felix (Anthropic Team Member)
How am I involved in this, I've been at Anthropic for a little bit, but this is the product that my team has built here. We've sprinted at this for the last week and a half. What we're trying.
- [36:47]
Dan Shipper
The last week and a half, that's it.
- [36:49]
Kieran
Come on.
- [36:51]
Felix (Anthropic Team Member)
To be clear, I think many people have had the idea that something like cloud code for non coding work would be helpful and useful to people. And fundamentally what we're going to do here is we do want to help people out with their work, like whether that's a personal thing or a corporate thing. And we've had a different number of prototypes and particularly before Christmas. But I think over the holidays, one thing we have seen, I'm sure many people have seen this, is that an increasing number of people is using cloud code for almost anything, just like we are. Right. We're sort of like automating our entire lives with cloud code. So we were thinking what is a small early thing that we can try out and ship to people and iterate with them together to really figure out what is the right user experience, what is the right thing we're aiming to build and this is it. This is the sort of like research preview, very early alpha, a lot of rough edges as you've already seen. Right. There's a lot of things about it that I think we're going to improve very quickly, but this is our attempt to build in the open and work
- [37:53]
Dan Shipper
together with people out there. I love it. Tell us about some of the design decisions you made. An early one, for example, is there's a third tab instead of maybe adding a cowork mode into the chat tab, like how did you think about and what was the process to come to the design that you have currently for how the product works?
- [38:13]
Felix (Anthropic Team Member)
That's a great question. So I think one belief I have is that the current user interface that you see across agentic applications, not just in anthropic but across the industry, is probably going to change pretty dramatically in about a year or two. Right now we have these hyper specialized individual input fields and we have a lot of custom scaffolding around this specific task that you're going to do. But as we see the intelligence of models improve and as we also like maybe holistically as an industry, figure out a little bit of the generalization problem, I expect that we're actually going to see a smaller number of interfaces for a wider range of use cases. For now, what we're doing is the reason we broke it out is because we want to be Pretty transparent that this separate thing is a construction site that we're sort of letting you into our kitchen. We want to work together with you. We want to ship almost every single day, some new features, some bug fixes, try out some things. So this separate tab is fairly experimental, you could say on the Frontier or the Bleeding Edge, but it's just a little bit less polished and a little faster pace. And that's one of the main reasons why in a separate tab. There are some technical reasons too, I could tell you. One of them is that currently this is running on your computer. So your chats are local, they're not shared with other devices. We're being being a little bit more aggressive in how many agentic abilities for gift cloud. Those are the main reasons.
- [39:36]
Dan Shipper
How did you think about. Because I feel like that's such a huge UX hurdle to get over. How did you think about letting people know, hey, this is actually running on your computer versus chat, which is in the same application as not? Yeah, that seems so hard.
- [39:50]
Felix (Anthropic Team Member)
Yeah. I think the dream that I have, and I'm sure many people have this, the dream that I have is that it doesn't really matter. Right. Like, where your code runs should be technical implementation, detail, and it should matter to people as much as when you visit the newyorktimes.com like, is it using websockets or not? And it's like, who cares? Yeah. I think for us right now it's an opportunity to move a little bit faster and to ship a little bit quicker and also like, work a little bit closer with the people for whom we're building this. I have this strong belief that it's very hard to figure out a great product in isolation by yourself.
- [40:23]
Dan Shipper
Right.
- [40:23]
Felix (Anthropic Team Member)
You sort of like go up into a cave and you work on something for a year and eventually comes out. I think it's really hard to build a good product that way. And I often like to remind people that like, even the first iPhone was missing. A bunch of things that we sort of consider to be table stakes. So, yeah, I think it's a pretty big hurdle, but we're okay with that for now because we do want people who, who are signing up for this. Right. To like, sign up for it fairly intentionally.
- [40:52]
Dan Shipper
Yeah, I think that's a really interesting pattern, is like, let's ship really fast and we'll ship it as a new thing in the app that maybe fewer people will click on so that we can get it out in the open and start iterating together, rather than like, try to make it perfect. Especially in this world where it says, you said you were working on this for a week and a half. This version for a week and a half, which is kind of insane. Um, Kieran, do you have any questions?
- [41:16]
Kieran
Yeah, I'm. I'm curious. Like, clearly this is the version that's out now. But, like, what is the version, like, in Yalls heads? Like, what are the. Like, where do you want to go next? Or like, what are the things you're dreaming about? You use the word dream. Like, what are those things where you want to. Want to go? What is. Because I'm sure everyone on the team had like wild ideas and then you were like, no, we need to ship Monday. So let's just like, what are. If you can share any of those, we'd love to hear those.
- [41:50]
Felix (Anthropic Team Member)
I love that question so much because I think I actually have the same question for the two of you, right, which is, where do you want this to go? What do you want to do? I've already heard you say you kind of want to give it to access to your entire computer. The multiple choice thing where you're like, actually, can we shift around a little bit on how we want to do this? I think right now I am much more in the mode of, okay, let's see what people think and then try out a billion things. Some of them will probably be the wrong thing, some of them will be the right thing. But I think it's much more interesting to me what people want to do with this rather than like, what's my own personal dream or vision. And in the things that I've sort of built in the past, this was always. This was always the thing that happened. Right? You have like an idea of how people will use the thing that you build. They actually find a use for it in all these other ways and then you lean into that. So I'm really hoping that. I'm really hoping that we can learn a lot about what do people want? What do people not want? What do they like? What do they dislike? I'm sure people will dislike a few things about this and then we adjust and iterate on it.
- [42:49]
Dan Shipper
That's the really cool thing about Go for it.
- [42:52]
Kieran
Yeah. So Boris is very good in building cloth code in a way that people can figure out what they want. Is there a way. Do you use that strategy in a way as well, where you give some building blocks or things for us? Like, for example, can I include my own plugins or skills? Or like, is there a way for people to experiment Inside Cowork as well in the, like, maybe the non coding way. Or is it really like, this is the product? That's what it is. Like how. Because there's a cool balance between like how cloth code works and people that use it because it's super hackable. Like, is there a similar philosophy in Cowork as well for non coders?
- [43:38]
Felix (Anthropic Team Member)
Yeah, like, very composable. Right. Like the first thing you said about, like, Boris being very good at steering cloud code in this direction of shipping early and then iterating on it and seeing how people use it. It's really funny that you mentioned that because I think one of the reasons we've shipped today, maybe shipped a little earlier, was because Boris pushed me and was like, hey, you should probably show this to people, see what they do. And on the composable piece. I think the thing that I found most impressive in my own work over the last couple of weeks and maybe sort of the last two months, is that I'm really leaning into skills. So instead of like previously writing MCP tools like this, like very specific harness that is like very tailored towards just Claude, I instead just write Skills. Sometimes I still write a binary and then I describe in a skill how to do something right. I'm like, what's a good example? I'm working on like a marathon training plan for myself and I wrote a little binary that fetches all my athletic activities from various pages. But then I just write in markdown in a skill file. Hey, Claude, if you want to make a training plan, like, please follow the following guidelines. We do automatically load any skill you have installed in Claude AI into cowork. And I think that's probably going to be increasingly, especially as models get smarter and especially with Opus 4.5. It is so good at following skills. So Skills is probably the primary hackable surface that I'm exploiting right now.
- [45:02]
Dan Shipper
That's great. One thing you said earlier in the conversation is you think that there's going to be fewer UI services. Does that mean that over time there'll be fewer UI services? Does that mean. Because there's a lot of debate over the last couple years about is chat the final form factor for AI? And everyone was like, no, we need more ui. Are you. Is that. Are you putting your stake in the ground as natural language actually is here to stay and we're going to have fewer UI services where you just talk to an agent, maybe an agent orchestrator that goes and talks to a bunch of other agents. And that's the kind of form factor you're, you're pushing towards. So it looks a little bit like how cloud code does today.
- [45:41]
Felix (Anthropic Team Member)
Yeah. I think this is still very heavily debated and there's certainly no anthropic viewpoint. I'm not even sure that there's a viewpoint that my friend, fairly small team would like holistically agree with.
- [45:51]
Dan Shipper
Right.
- [45:51]
Felix (Anthropic Team Member)
I think people have very different visions about how will people interact with AI and models in the future. If you ask me very personally, I think, I believe two things. One is that the chat input and its various forms, not just for models but in general. Right. Like the idea of this text box and you put into the text box what you want. If you generalize it enough to say even google.com or the ad address bar in Chrome is like a I want something input box. I think that is going to stick around for much longer than we all think. That's the first thing I think. I think we will continue to have something that looks like a search I want something box. The second question is like how many separate boxes do you have?
- [46:33]
Dan Shipper
Right.
- [46:33]
Felix (Anthropic Team Member)
Like do you have one box for code? Do you have another box for maybe like your personal attainment? Do you have another box for healthcare related concerns? I'm not sure we're going to have too many boxes of those. And there too maybe I would go back to Google. I think I sort of remember the early 2000s where you did different search box for every single Google sub product. And increasingly you just type what you want into your Chrome search bar.
- [46:57]
Dan Shipper
And yeah.
- [46:58]
Felix (Anthropic Team Member)
You no longer actually go to sub page of sub page. I'm in the mode right now of looking specifically for a shopping thing. So I go to Google shopping.
- [47:07]
Dan Shipper
Yeah.
- [47:07]
Felix (Anthropic Team Member)
And I would be surprised if we don't see a similar generalization that is smarter about figuring out what you want to do in the future. We might still have different interfaces where it's sort of like splits out.
- [47:17]
Dan Shipper
Right.
- [47:17]
Felix (Anthropic Team Member)
And understand. Oh, I understand that you're trying to do X, therefore I'm going to show you UI for X. But the entrance point, I think the,
- [47:26]
Dan Shipper
yeah, the interesting counterpoint to that is something like Microsoft Excel, which I think it also has some similarities to the way that just generally AI works is it's this general purpose product. It's super simple to get started. You can make things endlessly complex with Excel and then Excel spawned the B2B SAS wave. Like you probably don't get B2B SAS without Excel. So I think there's also the other argument that you have these sort of general, really general tools and then people find power workflows within them that then get split out.
- [48:00]
Felix (Anthropic Team Member)
Yeah, yeah. I think Excel is like such a beautiful example for so many things because it's like for many developers something that sort of exists a little bit on the periphery.
- [48:07]
Dan Shipper
Right.
- [48:09]
Felix (Anthropic Team Member)
I've often heard the analogies between how many daily active users Excel has versus how many developers even exist on the planet. It's an interesting number. And I think the thing that I find interesting about Excel and the commitment it has from its power users is that those power users are not too interested in marginal productivity gains or marginal UI or UX gains over deep familiarity with the product. I think that's interesting. I think there's a lesson there in some shape or form, and I think I've actually seen that across various other surfaces where you as a developer sometimes look at someone's workflow and you say, oh, I can make this workflow slightly better for you. You if I make you a specific use case tool over here on the side. And then people sort of fail to adopt that thing because they're actually more comfortable doing specific things within their product. As an example, I think that's a lesson that I have learned. I was previously at Slack for many years and there's a lesson that I've learned there over and over again, is that you can make these separate surfaces that you think might serve people's use cases much better, but they will continue to just do it in chat.
- [49:22]
Dan Shipper
That's a really, really good lesson. I love that. Speaking of that, I think today is for the non developers, but I feel like there's a lot of developers who are watching this right now and you're someone who's, if you built this, so you're deep into how to build agent native applications. This is something that we've been thinking about and talking about a lot at every. We just published a guide called called Agent Native Architectures and we've been thinking about what are the core principles of agent native apps. And I'm really curious if these resonate with you, if you think they're wrong or if there are things that you would add that are part of how you guys at Anthropic think about building agents. So an example is parity. So one of the things that we think about when we build agents internally at every is whatever the user can do through the ui, the agent should be able to do. I see that a little bit. That's basically how cloud code works, but I see that a little bit. And what you built with Cowork where, for example, if you didn't pick the file picker, it'll automatically determine that you are asking it to pick a particular folder and it will do that for you without you having to touch the ui. So that would be an example of parity. Another one is granularity, which is basically, tools should be mostly at a lower level than features and the feature should live in the prompt or the skill so that you can combine tools in new ways that you didn't predict previously. And then that allows for the third one, which is composability, which is you can combine, you can combine those new ways and you get the fourth one, which is emergent capabilities. So people are just doing it for stuff that you didn't expect and you see the latent demand and then you build for that. This is essentially, I think, a lot of my summary of how Claude code works. I'm curious how this sounds to you. And if you think that we're missing anything or there are any things that you've learned from doing this in production at a huge scale that could make people better at building these kinds of
- [51:25]
Felix (Anthropic Team Member)
applications, I think this really resonates with me. I think one thing that's hidden in emerging capabilities is, is the inability, I think especially individuals and silo teams have to predict how an agent actually ends up being super useful if you give it fairly primitive tools. I think pushing down tools into like a general space is very powerful. Like the more composable they become, the more generalizable. The more generalizable the tool is, the more you will benefit from improvements on model intelligence. And I think for many developers that I've been talking to in the past, it seems like the rate at which model intelligence and models ability to call tools effectively improves is actually much faster than your ability to maybe churn out additional tools and educate users on them. And I think if you take a step back and you think, how can I build a very generalizable tool? You have a much better chance to build something that can adapt to new use cases. I think that resonates with me quite a bit.
- [52:28]
Dan Shipper
What about the trade offs? I've been talking to Kieran about the trade offs in tools. Kieran, do you want to talk about what you notice in Quora and what you're thinking about?
- [52:37]
Kieran
Yeah, I think putting things in a prompt is great and then having the tools, but we need to now suddenly create tools that then read skills or something like that. So we have to invent this meta layer skills is just in time. Prompt injection. Like we need to create that thing. And now everyone that's building stuff, unless you use cloth code or the cloth SDK, it's all built but like is this thing like now there's like this struggle of like oh, but like tools are that you can describe stuff in the tool or you create a tool that then wraps around it and then call something else. So there's like this friction there. And it is great to make things composable. If originally you create for example five tool calls, one a search email, read email, this and this and this. But you can also say no, we do just do an execute tool call and we create skills that can do those things or an MCP or some abstraction there. So there's this change happening and obviously this is the close code is and the cloud SDK is a very good push for that. But I feel friction there. I'm sure you felt that friction too. So maybe you have some best practices for people that are stuck in the old fashioned AI world and need to go to the more agent native things that you learned or that you've noticed because you've implemented on top of I assume Claude SDK maybe or some very like there is. So you use that and you implement things on top. So I'm very curious if you learned anything there.
- [54:25]
Felix (Anthropic Team Member)
I'm not sure that I have any like wisdom from the mountain that is going to be more valuable than yours. But I think what you're saying, I think what you're saying that resonates with me is that you sort of need to make a call, right? Like which part of the outputs do you want to be non deterministic and where are you comfortable? Where you're comfortable relying on model intelligence. And every single time you do rely on model intelligence, if you pick a cheaper model or like a dumber model and those parts also go down in quality and I like to break up my workflows into like the non deterministic and the repeatable parts. I think the more repeatable something is and the more, more easily I can say this will never change. And if you get smarter, I'm not going to benefit here at all. I think that's a place where it might make sense to write a tool. And in a sense we're already doing that, right? Like we're not implementing. You could give Claude a very generalizable write assembly code tool, right? We could be like just call GCC and figure out whatever you want. But we don't, right? We give it like a yes, right?
- [55:32]
Dan Shipper
Because that's dense ideal, that's the most granular you can get. Yeah,
- [55:38]
Felix (Anthropic Team Member)
but I will say, I will say that I think when I talk to developers out there, and depending on how sci fi you are, I think even that assumption is a little bit. I wouldn't bet too much money on it. That assumption is certainly under attack. Like, the idea of, like, should you give Claude any tool at all? Or should it just be like, claude, here's memory, Start writing ones and zeros, go wild. It's an interesting area. It's like hard for me to know right now.
- [56:07]
Kieran
No one knows, but you learn stuff. You created skills for exactly this reason because you needed more than just a slash command and a sub agent. And you were like, we need the clothmd to be better. But I guess that's why skills were created. And clearly that's working well. I resonate with you saying skills are amazing. This is also like, I'm creating skills every day and I love them. So clearly there's something there. But when do you not have it be a skill? It's very interesting to me.
- [56:48]
Felix (Anthropic Team Member)
I think this is such a fun conversation. Someone you should actually talk to at some point is Barry. Because Barry is the one who is at least inside the company. Sort of like our skills person is the person who essentially came up with skills. And for us, fundamentally, skills were a little bit of a byproduct of the same tension that you're describing. So what we wanted to do is we wanted to make a very easy way for people inside the company to get dashboards.
- [57:15]
Dan Shipper
Right?
- [57:15]
Felix (Anthropic Team Member)
And we have a. We use one of the popular data providers where we keep a lot of our data. And we were trying to figure out, okay, do we build really specific tools that fetch that data and then compress it down into like a specific format? The first couple of dashboards closet in the building looked a little, you know, this was before 4.5. The dashboards were not ideal. Every third or fourth dashboard it generated was like a little lackluster. So we did think about, okay, do we like super parameterize it and basically build like a fixed dashboard that you cloth and only plugs new data into. But in that process, while building that we discovered, hey, if you just tell Claude how to effectively query this data source, that it can use SQL and that it please follows the following design guidelines for making dashboards. Suddenly you get something very, very good and you get it very good every single time. And then you also give people. This is the emerging capabilities part. Then you also give people the opportunity to say, hey, Claude, I understand that you're following these principles for dashboards. But I also want, I don't know, different chart type or I want to combine it with other data and that's then where things really open up.
- [58:24]
Dan Shipper
Right. That is really interesting. I feel like the one. One way to talk about why you might want to skill instead of just having it have GCC and just everything is just in time is it's like about sharing something repeatable with other people that you can talk about. And there is something. Actually not everything should be just in time because you want to do the same thing over time with a group of people. And that's. That's kind of a skill, I guess.
- [58:52]
Felix (Anthropic Team Member)
Yeah. That's sort of how we operate too as humans. Right. Like when I join a company, someone tells me how to book a flight,
- [58:57]
Dan Shipper
like, yeah, yeah, yeah, yeah.
- [59:00]
Felix (Anthropic Team Member)
How do you get a room? I think a lot of us sort of operate even as humans on a long list of markdown files with stuff in it.
- [59:08]
Dan Shipper
Felix, I want to give you an option. You've been very generous with your time. I want to give you an option to. To hop off if you want. We would love to keep chatting. We have endless questions, but I'm sure you have a lot, a lot going on. Do you need to go or do you want to keep chatting?
- [59:22]
Felix (Anthropic Team Member)
I think this is a good time for me to bounce, but I'm not going to go. Before both of you give me one thing you would like us to change.
- [59:29]
Dan Shipper
My easy one is just YOLO access to my whole computer. Okay. Okay. And make it easier for me to know whether it's working on my computer or working in the cloud on a chat and make it easy for me to use it on mobile.
- [59:44]
Felix (Anthropic Team Member)
Okay.
- [59:45]
Dan Shipper
Yeah.
- [59:45]
Kieran
Plus one on the mobile. But my favorite thing would be the ability to add my plugins. So just my. I have a marketplace with plugins. I just want to hook it up to GitHub and.
- [59:57]
Felix (Anthropic Team Member)
Fair enough.
- [59:58]
Kieran
Yeah, yeah, like. Because now I'm like adding things in the app and then copying it there and like it's. They kind of like it. Probably I can just copy it somewhere, but like just native support for a marketplace and then adding it and syncing it. That would be absolutely great.
- [60:13]
Felix (Anthropic Team Member)
Thank you. I appreciate both of those quite a bit. We're going to take those back, going to tell the team about it and for everyone else on the chat, like find us on the Internet, send us what you think. We're quite interested in hearing from people and adjusting our roadmap.
- [60:27]
Dan Shipper
Thank you so much, Felix. Thanks for Building this for joining.
- [60:30]
Felix (Anthropic Team Member)
Thank you.
- [60:31]
Dan Shipper
Have a good one. All right. That was awesome. So cool. Thank you.
- [60:36]
Kieran
Yeah.
- [60:36]
Dan Shipper
And we have so many people on this Stream. There's almost 10,000 people here. Oh my God. If you're joining us for the first time, we just had Felix on. Felix is a member of the technical technical staff at Anthropic and he was talking to us about Claude Cowork. If you are looking at this stream then you probably know what Claude Cowork is. But I will share my screen and show you again in case you're. In case you're wondering. It is a new version of CLAUDE that's like Claude code for non technical people. It looks a little bit like this. We've been testing it at every where. The only subscription you need to stay at the edge of AI every to. We get access to this stuff before it comes out and we do vibe checks like this. We do them live. We also write them on every to. We got this earlier today, so we were just testing it out. Here's an example of a task I gave it and you'll notice it looks a lot like normal Claude, like the normal Claude chat. The differences are a bit subtle, but I think that they. It does make a big difference. So instead of a chat you have tasks. When you look at the, you know, for this query, or maybe like, let me find a better example of this, like this query for example, you'll see that Cowork ran for a really long time. If I asked the same query, I wanted to go to our every agent native guide and walk through it to do a UX review, Claude would have stopped after a couple turns and given me a pretty good answer. But this does a lot more research. So because it's working on your computer, it's going to be able to work for a long time. And I think that's sort of the key thing that you can take away from when you might want to use Cowork and how Cowork might work differently than regular Claude, which is for the first time, if you're non technical you can ask your computer to do something and walk away for a while and then you can also ask it to do many things in parallel. So this is an experience that if you're using Claude code you have a lot with programming. But I think most non technical people are still in the kind of like turn by turn era of using chat and just expecting a response almost immediately. And this is much more of an built to be an async experience for non coding tasks. Like data analysis or research or writing documents, all that kind of stuff. Like Felix said, they built this in a week and a half, which is crazy.
- [63:08]
Kieran
Which is the new normal, by the way. So anyone who's coming up with a PRD in two weeks. Nope, you ship the whole thing a week and a half?
- [63:16]
Dan Shipper
100%. There are some rough edges here, but they're going to be improving them really quickly. I think it's really cool. The pattern that he shared that I really think is interesting is I was asking why even add a new tab like a Cowork tab? And he said we essentially needed a playground to mess around and do stuff that is a little bit less polished. It's nice to have this extra tab, which is an interesting pattern in AI where now you can build so quickly. I think we need more patterns for what to do with that. Kieran, any reflections from our conversation with him or anything you're thinking about right now?
- [63:52]
Kieran
Yeah, I think it's like, why use Cowork over chat or code? I think that is always my first question. And I think if you're a non technical user, just think of Cowork as something that is chat. Just try Cowork as the new version of chat, the better version, or a different flavor and, and just open it and do the same things you've been doing in chat, but you see that it is different and you can do things like one really, really important thing is in chat, if it is responding, you cannot send a new message. When it does something, you're like, oh wait, wait, that's wrong. With Cowork you can cue a message and like while it's working say something new. So like just those tiny things are very, very handy. So just try Cowork instead of chat. It has skills like you can generate documents, Excel sheets, PowerPoint presentations, PDFs, like it can do all of those things. So if you are applying to jobs, like just upload everything and say, hey, can you rewrite my resume for this job? And like try it out like things you would manually do normally and also connect roam, you can enable it, maybe you can show that. How to do that then in Connections.
- [65:10]
Dan Shipper
Yeah, so if you go into your settings and go into, I guess is it connectors or is it. Yeah, it's in connectors. Yeah, Control Chrome. So you can add that in there and then it'll just be able to basically use Chrome on your computer, which is really cool. I mean, one of the things I had to do that is totally new and interesting is do I have It. I had it go through my Twitter feed, my X feed, and I asked it to just tell me what was hot right now, what are people talking about? And that's really cool. That's ordinarily something that you would have had to pay a lot of money for an API for. And this can just do that without really a problem, which I love. I love that. Oh, here it is.
- [65:57]
Kieran
And if you use Chrome and you're logged in on things in Chrome, it's already there, so you don't need to log in again. It just uses your Chrome itself.
- [66:06]
Dan Shipper
Yeah, exactly. And I use Atlas, so I had to log in all this stuff. But if you use Chrome, it's great. So, like, look, this, it just, it just read my Twitter feed. That's so cool. There's so many possibilities if you're a writer or a marketer or anyone that does research, especially research on things that don't have APIs that you can now do without much trouble. And to your point earlier, Kieran, like, when you're thinking about, okay, what would I even use this for? I think that one of the things that we try to embody at every is we try to be really curious and know that if you're trying new technology for the first time, the first five things you do probably aren't going to work. But there's like a really. There's something really fun about being curious right now, because Felix, for example, the guy who made this, doesn't even know how we're going to use it. He needs people like us, anyone who's watching this stream, to mess around with it and figure out all the new, interesting, emergent ways that it could be used so that he can make the product better. So it's the first time where the software developer is more like setting up a playground, but has no idea how the playground is going to be used. And the users are the ones that are being creative and figuring out things to do. And so if you approach this with curiosity, it's like a huge opportunity over the next couple of weeks to figure out what this is for. I think another important thing is people tend to have this thing that I like to call capability blindness, which is, I tried this once before, three months ago. It's not never going to work with AI. And the really interesting thing about AI is, like, it changes every couple months. Like I freaking, like Opus 4. 5 just totally changed everything. Stuff that had never worked before started working now. And so if there's something that you really want AI to do, like, for me that has always been. I wanted to do copy edits. Ooh, we should see if it can do copy edits. I'm going to set that up. Um, so I've always wanted to do copy edits, and every time something new comes out, I just try it because I know at some point in the next couple months it's going to start working. So I'm going to actually set that up. I think that'll be a really fun demo. Kieran, do you want to. You want to talk about what? Anything on your mind so that I can take a little bit of time and make the Make. Get the copy edit set up. Yes.
- [68:30]
Kieran
I'll share mine here. I'll just share my screen here. So. Yeah, if you share my screen.
- [68:39]
Dan Shipper
Oh, sorry, yeah. Forgot that you were. That I was. You could see my screen. Okay.
- [68:45]
Kieran
We're professionals, everyone.
- [68:46]
Dan Shipper
Yes, obviously.
- [68:48]
Kieran
We started live streaming last week. Give us some slack. Yeah, cool. Yeah. So we talked about skills a lot. You might think, like, what the hell is a skill? Isn't that just a prompt? Yeah, it's a prompt, but it's also more. It's more like, yeah. And what can you do with it? So this is the most hackable way for cowork. So if you want to personalize your coworker experience, this is the way Anthropic says. So I asked, what are all the skills I have and how can I use them? So, for example, these are document creation skills and it just learns how to create Excel sheets. And these are some of my own. I love Swiss design, so I have like a Swiss design scale, a Gemini image gen where you can get nano banana images inside cold coat. I have a DHH Ruby style to roast my coat. And yeah, all these things. So I just create scales for everything. So I have a 3D print scale where I needed to print some 3D things. I was like, I'm sure cold codes can do this. And so I created a skill and these skills. Normally what I do is I go to chat and say, Can you deep research how Dan Shipper writes? So, for example, if I like Dan's writing, I might do this.
- [70:21]
Dan Shipper
And actually, I don't know. That guy's kind of a blowhard.
- [70:25]
Kieran
I know. So normally what I do is I enable deep research here and go hard and it runs. Deep research runs for like an hour sometimes, which is great. And then I say, can you create a skill for this? So I'll just fake do this, but then I create a skill for this. And there is this thing in Claude, in Capabilities that you can enable, which is really cool. The example skills. So you go to capabilities, example skills. And there's a skill creator from Anthropic. If you enable this, after you did the deep research or anything you can say, create a skill out of this, obviously you can do that here as well. So you can do research in Cowork and create a skill out of this. And it will be loaded inside, inside here. So for example, now KU design with Swiss design a very beautiful chair for me and create a STL so that I can 3D print this miniature 4cm high. So let's see. So I do this and it shoot them. KU design. Okay, so hopefully it's. Yeah, so it's loading the two skills here. You can see the context it pulled in the Swiss design skill and the 3D print skill. So I like that that you can really clearly see this. And what it does, the skill will just inject a prompt where it says just do these things. Make it look good. Don't use inter or whatever. This is the aesthetics. So it's now doing things. And there is also in skills, you can put like. Like scripts, Python scripts, whatever script you want, binary scripts, anything you want to run. So if you want something programmatic, you can encapsulate that into a script and put it in a skill. So every time it runs, for example, you want to check if it's the shape or if it's following a certain static or if it's actually linting. Well, you can encapsulate all that in a skill and trigger it as well. So it's now doing this, which is really cool. And the skill will create a STL file which I can then 3D print.
- [72:50]
Dan Shipper
So we'll see.
- [72:50]
Kieran
Like, a chair obviously is maybe a little bit hard, but why not? So this is how you can create your version of Cowork with skills. So you can capture and find the things you do and encapsulate your style and everything you do into a skill and then have it available here. What is cool here is on the right side, you can see the STLs out. The preview. There's a preview. Let's look at the preview. Okay, so it's a little bit hacky still because. Or here, the sidebar. There we go. Okay, this is the chair. Beautiful. So this is SVG. It's not 3D. Let's see. Can we look at this? No, we cannot look at at this. But yeah, we have a stl. Let's see if I can open this. Opening Bamboo Studio. I'll See if this looks any good. I'll share my other screen. Share screen. And then we go to them. So here is my Swiss design chair.
- [74:01]
Dan Shipper
Amazing. We need the skill. Kieran, give us the skill.
- [74:05]
Kieran
Oh, yeah. So the, the actually, the 3D skill is. Is on my machine, but I'll share it. I'll push it to the. To the interwebs. And that was actually my request to. Yeah to Felix is like, can I automatically like pull these things into my. My coding experience or like in Cowork that I. If I push them online? So, yeah, I will share this. But it's. Yeah, make your own things go wild. And it's kind of funny to then print this. I'll print it and take a picture later.
- [74:39]
Dan Shipper
I love it. So I. If you've just joined one of the things that we're talking about with Claude Cowork, I think that one of my bars for AGI is can it do copy edits in a Google Doc? Like, it's surprisingly hard to get these things to do that. So what we have at every. Obviously we publish every single day. We have a very high bar for the copy that we publish. We want to make sure everything is like really, really clean and beautiful. And so we have a skill that's in Claude that is our every proofreader skill. And what I wanted, what I wanted to do is see, can Claude Cowork copy edit a Google Doc? So this is the Google Doc. This is an article that we published last week by Katie Parrott, who's one of our writers who's fantastic. She's going to be the one writing the vibe check on Cowork for later today. So look out for that on every. Every to. So I just said like, okay, I have an every proofreader file. I just downloaded this go file and gave it. Gave it to it in Documents. Can you go to this Google Doc and make edits as suggest changes? And then this is actually really interesting. So, like, it's just hard for AI to do this because it has to go through the copy editor and just look for every. For each rule in the copy editor. It has to go through every part of the document and find all the violations. And that's just really hard to do. But you can see I said go do this. This is not something that I don't think that the regular cloud app could really do this. Maybe Claude, if used in Chrome, could, but other than that, it could not do it. You can see it loaded the document. It's getting all the text. It scrolled through the whole document and it's now clicking on. It looks like it's clicking on Suggesting. Let me see if I could actually find the Chrome tab that it has opened. Yeah, you can see that it's. Here's what it's doing. It successfully got itself into suggesting mode and now what it looks like is it's searching using Find and Replace for errors that it found and we'll see. Now I need to double click on Editor in Chief to select it and type the replacement. Let me triple click near it to select and manually make the edit. This is so interesting. I feel like watching. This is like watching a video game. I'm like, oh, yeah, come on, Claude.
- [76:58]
Kieran
It's really entertaining.
- [77:00]
Dan Shipper
Yeah. But Google Docs is like the final boss of stuff for AIs to use because it's just like. So it's such. Ultimately such a simple application, but it's the way that they built it is so complicated. It's like not actually real HTML. It's like a whole. It's just. Just really hard. So, yeah, it's just struggling. So if you're at Anthropic and you're watching this, if you could please improve your computer use to actually be able to use Google Docs, well, that would like totally change my life and change the life of Kate, our editor in chief and a bunch of other people here. So, yeah, please, please do that. And also everything about AI is like a video game. Absolutely true, Kieran. Go for it.
- [77:44]
Kieran
Yeah. So I'm thinking, so we have this skill, but is it like, skills can be maybe optimized now to actually make use of sub agents and things like that that maybe were never available? So it might be also time to rewrite our skills a little bit. If you created Skills for Claude, maybe we can push it to use sub agents or you execute scripts more or like do some things in a programmatic way. So there might be an opportunity also to do more of that.
- [78:16]
Dan Shipper
Yeah. Katy Parrott, who wrote this article, says surely this is the cleanest copy that has ever copied. It's true. It probably hasn't found many errors because, Katie, you didn't make any. Um, but we'll. We'll see. It's. It's still working on Editor in Chief with dashes.
- [78:32]
Kieran
It's doing something.
- [78:33]
Dan Shipper
Thank God this is working. This is able to work Async, because if I had to, like we were actually, you know, looking over the. Over its shoulder for forever, it would, it would not be particularly useful.
- [78:43]
Kieran
Yeah, we'll do a live Stream or Claude should do live streams of it doing work.
- [78:48]
Dan Shipper
Yeah, yeah, Actually that'd be really fun. Claude doing work.
- [78:51]
Kieran
Tv and I see it mess up. Yeah.
- [78:55]
Dan Shipper
Here, I'm going to give it a hint. Hey, buddy, you can just type in there and replace it. You don't have to use find and replace. Let's see if it. This is.
- [79:06]
Kieran
Yeah, yeah, you sold. Do that.
- [79:09]
Dan Shipper
Yeah. This is another thing that. Like that this does that is different from if you're using the chat, which is you can see this button says queue, add to queue. And this is going to be a lot. The UX of this is a lot more like the UX in cloud code, where you can just send messages. Even if it's working and it'll deal with the messages, you don't have to wait for it to respond to your last one. And I think that's really, really good and important for Async conversations. And we got our first compact. The bane of every Claude app user's existence is compact. So we'll see. Coming up after the break, we'll see if Claude. We'll see if Clyde can replace Editor
- [79:54]
Kieran
in Chief with Dash or forgot what it was doing and does something completely different. We'll see. So one thing, it says Q, which is a little bit confusing because Q in my head would mean like, do this after you finish the one before. But it's actually a little bit smarter if you say, stop, stop, stop what you're doing. This is terrible. It will actually look at the text and think like, hey, do I need to change anything what I'm doing right now? So it does either pick it up immediately or if it's like, after this, do this. It will actually queue it. So there is a little bit more flexible than just like injecting it. It will look at it whenever you add it, which is good to understand as well.
- [80:37]
Dan Shipper
I wonder if Claude is getting performance anxiety because it knows that 13,000 people are watching it. Okay, dude.
- [80:45]
Kieran
Oh, I thought there was.
- [80:46]
Dan Shipper
13 people are watching this. Please just type. You're so close. If you want more extremely smart and insightful takes on AI, you should subscribe to every. Every is the only subscription you need to stay at the edge of AI. It's Every to. We have a pretty cool business. We do ideas, apps and training at the edge of AI. So on the idea side, we have a daily newsletter we get our hands on. We get our hands on stuff like Claude Cowork early on the day that these things come out. We do vibe checks which tell you from Our team as we're using these in our day to day life and work, what is it good for, what is not good for? That happens for products, it happens for new models. So like when Opus 4.5 came out, we had a review on the day of. I can show you that we also develop apps ourselves. I'll talk about the apps in a second. But this is the article that we wrote on the day that Opus 4.5 came out. This is by Katie. Again, Katie. This is. Katie's article is by Kieran and by me. If you've been seeing all of the hype about Claude code and Opus 4.5, we started talking about it November 24, 2025, the day it came out, and we said it's the coding model we've been waiting for. It took people about a month to catch up to that. If you really, really want to know, if you really want to know what's going on at the edge of AI, it's really good. I'm biased, but I think this is a really good read. And our vibe checks are really good. We also have apps. We have four apps that we build as part of every. We have one called Quora which, which, which Kieran builds, which is an AI email assistant, CORA computer. We have one called Monologue, which is a speech to text app. You can see Monologue right here. This is Monologue I'm talking and it will type in here. I don't want to mess up our buddy Claude, but it's sort of like Whisper Flow or Super Whisper and it copied it to the keyboard. You see it? It just pasted my text in there. We've got a couple other ones. Spiral, which is an agentic ghostwriter, and Sparkle, which is a file manager file cleaner. It's all available for one subscription. You pay one price and you get access to all the ideas. So everything that we write, all the apps. So for AI apps, AI apps at the edge of AI and training. So we have camps that you can go to. We're doing a Cursor camp where the team from Cursor is joining us in a couple of days. And you get to learn directly from the Cursor team. You saw we had Felix from Anthropic on earlier. We do these all the time. We basically teach you how we build. Everyone at Every is a builder and writer and as we're using these tools like Cloud Cowork and Cloud Code to build apps and do writing and do design and all that kind of stuff, we bring you along for the ride. You should subscribe every to. And we're still on this extremely scintillating view of Claude trying to make suggested changes in a Google Doc and I think we can call this one just we need a copy edit Google Doc copyedit benchmark and it has failed. That benchmark is not saturated. We're still not AGI folks, so stay tuned for hopefully the next model. Oh, Anthony. Anthony Morris from Anthropic. Huge fan of Quora. I love to hear that. Kieran, does that make you feel good?
- [84:23]
Kieran
Yes. Yeah, we have a few anthropic peeps using it.
- [84:27]
Dan Shipper
That's pretty great. Thank you. Yeah. Kieran, what's on your mind?
- [84:34]
Kieran
I think on my mind is I want to try this out. This makes me very excited. Like this, this feels like something that was missing. Even as a coder, I want to use this. So, like, if you are using cloth code, probably this is easier for you to use and I really want to just see what it can do, how we can use it in our daily lives. I would love iOS integration, scheduling things, pushing things to my phone, like me chatting with it. And also one thing that looks like it's not super present now is storage. So currently you can store or connect it to your local computer. But what if you're on your phone? Is there some persistence? Like there are projects in the chat side, but there are no projects in the cowork site. So like, I'm very curious how to use this, but what I love is it hooks into the skills and things I already have. So what I'm going to do is just whenever I would have started, started up a chat window, I'm now going to start up a cowork window and just see how it behaves and goes and just go from there.
- [85:44]
Dan Shipper
Yeah, I think that's a good. I think that's a good one. Okay, so we're going to have to write the vibe check in a few and probably also do some other work. I am curious if you had to give this a rating. So when we do vibe checks, we have red, yellow, green and then we have a gold metal for paradigm shifting. So when we, when we did Opus four. Five, I think you and I were both at the paradigm shift level for comparison. I think for GPT5, I was a yellow green and you were a yellow, I think when it first launched. Right. So where do you put this? If you had to give it a red, yellow, green or a gold medal?
- [86:23]
Kieran
I would rate it from just playing with it, like the UI and execution. I would say Yellow because it's kind of janky, but it's very interesting. We can say what we want, but it is kind of janky. But that's what they said. We made this to try out and I see Anthropic is very, very good at listening to people. I'm sure someone on the team here saw me click something or you do something. They're already pushing changes to this because that's how fast they go. But from like, from an idea, I would say green because I really think we should experiment more with the interface, what it is, and like giving this cold code moment to more people, like having more people that do normal work also starts to feel a paradigm shift of like Async work and really handing something to an agent. And I think this could be that because I don't see any other company do it like this. And obviously Claude is very. Yeah, like, it's a very good harness. So like let's milk it, let's make it better. But I would say ID green execution today, right now, yellow.
- [87:46]
Dan Shipper
Yeah, I think that's spot on. But to your point, Anthony Morris just said I have a PR up already from something Dan said. So. Yes. Lol.
- [87:56]
Kieran
Exactly. Yeah, that's why we do this.
- [88:01]
Dan Shipper
Welcome to the Anthropic product meeting happening live on X with your friendly neighborhood product testers. And yeah, I think that's totally right. They're moving super fast. I love how, I love that there's already going to be changes in the product from this. So for anyone who's watching or listening, you should give more feedback, do it in the chat, send it to people on X. They really actually do iterate pretty quick and I think that it's fine that this is a yellow. And the way that I think about it is obviously there's a cloud code moment on X over the last couple of weeks and that was the thing that they immediately just shipped something like how many, how many founders were like building a cloud code for. We were calling it cloud code Easy mode because it was an ideal. We were batting around internally at every two. And, and they just built it. They just went ahead and built it. And I think that's. That's so cool and it shows. They really have their finger on the pulse of what people want. That's, that's the thing that I think is good about a lot of anthropic stuff is you can tell they use it themselves and they're using it in a very AI native and agent native way.
- [89:10]
Kieran
Absolutely. Yeah, they get it and. But they're also listening. Like they get it but they also leave out things because they want to listen as well. They're not like this is the way, they're like, yeah, this is a way. And we're listening very carefully. And then they make iterations, which is great.
- [89:27]
Dan Shipper
Yeah. Speaking of which, this is a way that if you are a, if you want to build an AI, this is a way to think about doing it. We have a guide up on every about agent native architectures. So if you're a developer, this is going to be really good for a developer. It's also totally good if you're a non developer. You just click read with Claude or read with ChatGPT and it will tell you about what this is. But this basically boils down a lot of principles that we've kind of intuited from the way that Anthropic builds cloud code and now cloud coworkers and turns it into a really easy guide to think about. How do you build software in this era where agents are at the core of software? Instead of software being this sort of like deterministic thing that is built with deterministic code that are rules that are laid out beforehand by a programmer. Instead the core of something like cloud code or cloud cowork is just an agent and features are really just prompts to that agent to get work done. And that opens up a whole new territory of software to build and it opens up who gets to build software. So if you're a non technical person and you, you're feeling like, oh, I can't really build software, I promise you you can. You should really try cloud code. You should really try just a cloud app or now it's worth trying cloud cowork. I think that, that it could be really good for vibe coding stuff. And if you're thinking about how to structure your apps, it's actually kind of not intuitive because it's a whole new world for, for the way that you do programming. And so a lot of the, a lot of program intuition I think is outdated and a lot of therefore a lot of AI intuition about how to build software is outdated. So you need guides like these to help push your AI to do the right thing and build that, build the thing in the right way. I think we're, we're getting close to time. I'm going to need to hop off, we're both going to need to hop off and get some actual work done. This is fantastic. It's so fun where we do vibe checks for every new thing that come out we're doing this. These live streams as a new thing. Usually we just write them. But you should expect next time there's a new model drop or a new product release that we will have a live stream vibe check with our internal testing. Remember, we get all this stuff before it comes out and we'll tell you what we like and what we don't like. Kieran, any final words before we head off the stream?
- [91:43]
Kieran
No. Cheers. Thank you, everyone.
- [91:47]
Dan Shipper
Thank you.
- [91:48]
Kieran
Check it out.
- [91:49]
Dan Shipper
Check out every try. Clyde. See ya.
- [92:01]
Podcast Host/Announcer
Oh, my gosh, folks, you absolutely, positively have to smash that, like, button and subscribe to AI and I. Why? Because this show is the epitome of awesomeness. It's like finding a treasure chest in your backyard, but instead of gold, it's filled with pure, unadulterated knowledge. Bombs About ChatGPT Every episode is a roller coaster of emotions, insights and laughter that will leave you on the edge of your seat, craving for more. It's not just a show, it's a journey into the future with Dan Shipper as the captain of the spaceship. So do yourself a favor. Hit like Smash, subscribe and strap in for the ride of your life. And now, without any further ado, let me just say, Dan, I'm absolutely, hopelessly in love with you.