Claude Cowork GTM Workshop w/ Anthropic and WorkOS

Test test test.

Everybody, everyone in the back, can you guys hear me?

Test test test.

Can you hear? Okay. No, I don't think it's on. My stuff's on.

it's on. My stuff's on.

Test test test.

Test test test.

Test test test.

Use a handheld. Okay. Can everybody hear me here? All right. Cool. All right. Uh,

me here? All right. Cool. All right. Uh,

thanks so much for coming. Um, this is hands-on co-work. And we're going to

hands-on co-work. And we're going to literally spend an hour hands- on keyboard. So, I highly recommend taking

keyboard. So, I highly recommend taking out your computer if you've got co-work.

And I'm also going to show you a really cool um voice flow. But first, uh, work OS is the way that I describe work OS is imagine I'm a developer. Imagine my

developer friend and I have finally figured out the exact killer app that we wanted to build forever. We're going to do it this weekend. And we get it working. It looks great. We take it on

working. It looks great. We take it on Monday to do the first demo and it's a large company. And they say, "Do you

large company. And they say, "Do you have skim provisioning? Do you have audit logs so we can make sure nothing went wrong or there's no like access issues? Are you fully compliant?" Right?

issues? Are you fully compliant?" Right?

Um can you offer like social login, social SSO? Right? And we know as

social SSO? Right? And we know as developers that we could build that stuff, but we really don't want to maintain it long term because there's constantly, you know, issues with being discovered with SAML. We don't want to

patch those ourselves. So instead, we go to work OS and we we drop in um all of those features via JavaScript and then our app is ready to go up market and we can start chasing bigger deals and

actually sell it. So uh that's works in a nutshell and now we're going to get into the presentation. So the main uh bright line that I want to kind of carry through this whole talk is if you can do

it today, do it today. And we're going to see um exactly why. So I'm Zach. Nice

to meet everybody. I uh work on applied AI at work OS, but I also use Claude every day for everything. And I've been doing this before cloud code. Um and I use it for coding, writing, consulting,

finances, getting organized. Um, I've

even given it um the ability to read my biometrics so that when I talk to it, uh, I've prompted it to to push back on me if I slept only three hours that that last night or if I'm running a fever,

it'll help me plan and say maybe a nap in the afternoon and some medicine. Um,

but that's been kind of a fun experiment. Uh, another thing I love

experiment. Uh, another thing I love about this kind of agentic phase that we're in now is that I can still get a ton of work done, uh, maybe more than ever before, but I don't have to be at a desk as often. So I've taken to even

going on long walks doing a lot of my thinking and planning and architecture work kind of just in headphones and talking and I can even from the woods where I get a s a sell signal I can still look and and poke my agents and

get them to you know update PRs and do something differently essentially and that's been one of the really exciting kind of liberating things about uh this era of software development. Uh, I even

had Claude build me a watch app for my uh to that purpose to kind of continue to free me and let me just go wherever and keep working. And uh I will say that the hardest part of doing this was upgrading my Mac to get to the version

that I could install Xcode 26.2 which has cloud installed because once I was able to sign into cloud Xcode, it took about two and a half hours of iterating, testing, sideloading it directly on my on my watch and then it worked. And so

I'll show you guys just a quick uh demo here. I don't know if you can hear it,

here. I don't know if you can hear it, but essentially this is an app called Handwave. It runs on the on my um on my

Handwave. It runs on the on my um on my watch and I tap the button to find a Mac. It uses the Bonjour protocol.

Mac. It uses the Bonjour protocol.

There's a server, a bridge server running on my Mac where all of my Cloud Code sessions are. And so it can, let me go back here. Sorry. Um it can pull up

and find every single one of those sessions. And the cool thing is that I

sessions. And the cool thing is that I can then talk to them and say, "What was the latest response?" It can record my voice and send it. and it uses the entropic agent SDK to resume that session and then you know tell me what's

going on or I can give it a direction to continue. So you can imagine with um

continue. So you can imagine with um with this I can kind of wander around the um the house and continue to make progress and check on my cloud code sessions. So I just uh shared all that

sessions. So I just uh shared all that to say that I have found cloud to be incredibly capable for a wide variety of use tasks and that's why I'm really excited to share this uh with you all

today. Uh but first I want to just

today. Uh but first I want to just quickly demo something that compounds everything that we're going to look at today and I highly highly recommend it.

A bunch of us at work have been discovering this and finding it incredibly valuable and that is to essentially stop touching the keyboard.

So I would say for the last year or so now I have really not typed. I'm I'm

only typing things I don't want people to see when I'm in a meeting. Um and the rest of the time I'm really talking and speaking software into being uh documents, plans, architecture for whatever I'm working on. Uh, Whisper

Flow is the tool that I use, but there's a ton of other options. And um, I will say that I I thought I was a pretty decent typist at 90 words per minute as a developer, but I'm hitting 179 words per minute. And we'll see why that's

per minute. And we'll see why that's really powerful when you combine it with a Gentic Workflows. Um, the fun thing about Whisper Flow is I'll show you. You

can use it in two different ways. So,

let's just say I have a terminal here.

And if I hold down just this, then I'm speaking into a terminal. And I let go and there's the text. But what I really love to do is double tap it and then just walk around and free myself and

wander and pace the room. There's a

buffer that goes up to six minutes and I I literally rant like a crazy person and I just get all of my thoughts out from what I'm working on to architecture for this project to random idea I want you to save later um to oh by the way I also

have to do this errand in the afternoon.

Don't let me forget. And you know this is really a powerful thing and I just wanted to kind of share um a reason why.

So a couple reasons why. The first thing is when you're speaking like that and you have a tool like Whisper Flow that's using an advanced model to kind of translate what you're saying, um, you

get just kind of like cleaner text as a result. So all your ums and everything

result. So all your ums and everything and your hesitations get cleaned out and you just get the actual prompt. Uh,

that's pretty useful in itself.

But there's another thing here um, that's really serious and I can't unsee it now. So imagine that there's two

it now. So imagine that there's two developers sitting next to each other.

I'm using my voice and I'm running three different cursor windows or eight different cloud code terminal sessions, right? And then you have another

right? And then you have another developer is just still typing in their first prompt. But because I can speak so

first prompt. But because I can speak so quickly, so much more quickly than they can type, I could have three different agents already working in the background on their task while the other developer

is still typing things in. And if you imagine, you know, layering this on top of other really powerful agentic systems that are getting better and better every day like Codeex or Google Jewels, um

then you can see that there's like a way to kind of run through multiple sessions and parallel work tracks throughout the day if that works for you and then speak um all of your your you know desired

changes and get to the end goal significantly faster. Not twice like not

significantly faster. Not twice like not 2x faster, significantly faster.

Yeah.

Yes. Just a quick interjection if you need to connect to the Wi-Fi the password is enterprise explorer ready.

Thank you sir. Um another thing this is just to show exactly this kind of um uh the other thing I do and the brain dump is is incredibly useful is just you know come back from a walk and then dump

everything I'm thinking about right then to Claude. And uh again the value of

to Claude. And uh again the value of that is that you know my brain can be chaotic but Claude can take all of that and then quickly organize everything and tell help me prioritize and in a couple more turns of chat like I'm basically

like happy with the plan for the rest of the day and I can kind of be on my way.

So uh don't sleep on voice if you haven't tried it yet. I think that's um highly highly valuable and we'll see um I'll I'll do some of that while we're doing this. So with that said let's get

doing this. So with that said let's get co-working. Um this is the seat belt

co-working. Um this is the seat belt safety talk up front. So co-work is a developer that's in your machine. So it

can read, write, browse code, and run things. Um, the other way I like to say

things. Um, the other way I like to say is like inside your machine there are two developers. Like there's two wolves.

two developers. Like there's two wolves.

There's one that wants to constantly help you and make excellent changes. And

there's one that got slightly confused because you were, you know, it was writing in in multiple languages and has tons of file handles open and you said something ambiguous in chat in the middle of a cued message and it either

went and did something like delete stuff that should have stayed or it shared something that it shouldn't have shared.

And so this is the real uh core tension and danger right now. It's just that with the ability, you know, with the power, you also have the power to really hose things or yourself or your system.

So to that end, uh the way that I like to think about this is that speed requires safety and um when you're running an agent on your own device and not on a separate system, um I like to

reduce the blast radius as much as possible. So work in a specific folder

possible. So work in a specific folder uh especially initially as you're getting set up and working with learning co-work. Make sure there's nothing

co-work. Make sure there's nothing sensitive in it. If you're going to go do a task, create a new directory, put all the source materials in there, read the source materials yourself, verify there's no PII in there, verify there's

no API keys hanging out, for example, then you tell co-work it can work just in that directory and go go nuts. So,

you can see this is roughly the the difference, right? If you just go and

difference, right? If you just go and you're starting out and you point at a giant shared drive in in the cloud, um you could be in for a very very bad day.

Whereas if you make a single workshop folder and put only the data in there and the templates um you've reduced your blast radius. The other heristic that I

blast radius. The other heristic that I like to try and use for this is that um you know I'm driving this agent I'm telling it what to do. It doesn't wake up on its own in the morning and decide

what to do. So if I am hasty and sloppy and end up leaking something uh while using it, it wasn't Claude code that did that. It was me. Um, so that's the se

that. It was me. Um, so that's the se the the only, you know, safety and seat belt talk that we'll have. And this is the loop that we're going to experience tonight, which is uh the co-building loop. So when you're dealing with

loop. So when you're dealing with co-work or or other similar agentic uh systems, you decide what you want. You

tell it exactly what you want. You

describe it as best you can. It makes a plan that you can review or you could comment on. You can even talk to it and

comment on. You can even talk to it and steer it while it's working. It's going

to execute and build all those things and write scripts faster than I can. And

it's going to generally do pretty good work. and we can expect it's going to do

work. and we can expect it's going to do better and better work uh in the future.

And then you give feedback and you iterate and basically say, "Yeah, that's almost what I want. I hate these colors.

I hate that pixel art. What was this guy thinking? Like, change this." Right? And

thinking? Like, change this." Right? And

you're just constantly going through this loop until you get to the the actual end goal and the artifacts that you really want. And so, the way to think of this is that there's a ton of these new powerful tools that are coming

online. um they're powerful in their own

online. um they're powerful in their own right, but if you think carefully about the most optimal ways to combine them, you get something completely different than just using one of the tools. So,

when you add voice and you can now speak at like 179 words per minute and then you add an agent to or multiple agents that you can control and direct, then you add plugins to those agents, which

we'll see here in a second that Co-work ships with a ton of um you know already built-in functionality that's just common sense for all the types of workflows that we're doing. Uh and then you can start scheduling tasks after

you've gotten them working and defined.

And this is really where it's you're going to start to feel a significant lift. Um so plugins, let's look uh

lift. Um so plugins, let's look uh briefly at uh co-work plugins. Plugins

basically give uh co-work superpowers is the way I like to think of it. So if I open uh co-work here in any chat, let's open a new chat. And then anywhere here you can basically do the slash um and

you'll get this popup menu. I've already

installed sales and marketing, but if you go here to the ad plug-in interface, you'll see this is a ton of functionality that's already baked in and built and available for you to pull

into your workflows. The benefit of using these is that you do not need to constantly maintain all the prompts anymore. And the other benefit is that

anymore. And the other benefit is that as the skills improve on the back end and as we're learning more about plugins and they ship updates and improvements, you get those improvements. If I install the marketing plugin and my colleague Will installs the marketing plugin,

we're running the same commands, which is also incredibly powerful. And then

the kind of the final takeaway I want you guys to have is that at the end here, this really starts to get uh at another level even more powerful when you start writing to and reading from

shared team drives in a safe way. What

that's actually going to look like for each of you depends on um your actual business and and what you're doing. Um,

but I think that you can we can kind of start to see like the horizon with this stuff combining together. And it looks like um defining, you know, workflows constantly that can be automated, automating them away and scheduling them

and then producing artifacts at lightning speed that we weren't able to kind of consider before. So, um, that's plugins. And then the way that you would

plugins. And then the way that you would use a plugin is like let's imagine that I'm I'm looking to draft a blog post. I

do slash again. I start typing the the name of the plugin or the command. And

you can already kind of see in this popup the prompt what the prompt is kind of expecting and the parameters here.

And so um again just a faster way to kind of get at the functionality that you're looking for. So recommend

checking those out and installing plugins.

Um all right. So we've talked about plugins. Uh this is a good time to go

plugins. Uh this is a good time to go and install both sales and marketing plugin. Um and then we'll use them in a

plugin. Um and then we'll use them in a second here.

Any questions so far? Or Yeah. Can you

tell the password the Wi-Fi password?

Oh, I think it's enterprise ready.

Yeah. Yes, sir.

Oh, thanks.

Go ahead.

The same scale. So,

they're they're a little bit different.

My understanding now is that plugins in co-work are solely this is a separate ecosystem where plugins are already available in the co-work binary and you can say I want to install this or it's going out to like a marketplace that

they have. um it's not quite the same as

they have. um it's not quite the same as skills like skills are the markdown files where you you describe you know how what the skill does how it's used so that claude can very efficiently decide

if it needs to use that and then skills optionally include scripts at the end that the AI can run in order to get a deterministic output and I'll actually show you a cool custom skill at the end of this if we have time um I think

there's a lot of like bleed over in between uh plugins and skills in general right now but the the rough mental model I use for all of it is that we don't want to keep redefining finding this functionality on everybody's machine so

we all install the same thing and then you have the same issues with plug-in management you've ever had in software before which is versioning updates who's running what right the benefit of co-work too is that it could be less uh

of that pain if everyone's using the centralized plugins yeah good question any other questions or issues what you say you're normalizing terms of doing your brain build sort of voice to

cloud yeah um what I do is like if I find that I'm kind of just like overwhelmed. I

have too much going on. I can't really think clearly. I will just turn on the

think clearly. I will just turn on the the dictation mode and I'll just rant for six minutes and say like I'm frustrated. I really want to get this

frustrated. I really want to get this thing over the line, but here's the problems with it and I haven't yet figured out how I want to do this piece.

I'm also starving. I realize I hadn't eaten in four hours and at 5:00 p.m. I

got to remember to pick up the kids, right? And then, you know, it'll go and

right? And then, you know, it'll go and chunk for three minutes and it'll be like, you know, chill, you're fine. This

is what you're going to do. And I'm

like, yeah, that sounds reasonable, but I don't want to do that. It's like,

okay, do this instead. And now I'm happy with the schedule. Um, but that same process works for anything. Like it

works for just, you know, software projects. Like I can rant for six

projects. Like I can rant for six minutes about everything I hate about this architecture and how we got to improve it. And um, I find it just

improve it. And um, I find it just incredibly powerful because Claude can type six times faster than I can.

The the the cool thing about whisper flow at least is that it will go into um any curs wherever the cursor is on the screen. But then on in cloud, like it

screen. But then on in cloud, like it depends. If I'm using cloud desktop, it

depends. If I'm using cloud desktop, it would go in a single chat until the context window is filled or it's compacted. If I'm talking to cloud code,

compacted. If I'm talking to cloud code, I tend to just have multiple tabs open for each session for what I'm doing.

I'll tend to have like one task or one project per tab and then it's kind of saved that way. Yeah, great questions.

Anybody else? Yeah.

So, uh I've used clo a little bit and you know you got there's so many MCPS that each company is doing right now but it's so powerful.

Y would that be the equivalent? Would

plugins be the equivalent in cloud work?

And that's not my first question. And

second, have you seen most of the most popular tools like aio or like a CRM already doing plugins for glo?

Um, no I don't. So the sorry repeat your first question one more time.

Would plugins in cloud coord would be equivalent to last MCP servers? Yeah, the weird thing right

MCP servers? Yeah, the weird thing right now is at le I'm pretty sure I hope I'm not lying but the last couple of builds of co-work that I tried I wasn't able to just easily install MCP servers. So it's

more skills and plugins. They're a

little bit separate. But I do think in general in the ecosystem it's like MCP skills plugins are kind of in some ways similar. You could use them all in

similar. You could use them all in different ways and depending on your actual scenario and context one might be better but it's also I think kind of proliferating and everyone's trying to figure it out and it'll probably maybe

converge at some point is my expectation. Um a second question is I

expectation. Um a second question is I haven't seen it yet but we can look in a second there is an ability to in co-work to add um connectors here. So

right they have a ton. So what I have seen is uh like Slack and all those guys, granola, AWS marketplace, right?

Some of these that I haven't set up yet, GitHub. So all those would probably be

GitHub. So all those would probably be super powerful as well. Yeah, great

question.

Sorry. Yep. Hi.

Yeah, that do I need to be on max or I don't think so. I think it's honestly to be just candid, it's been pretty um buggy and una unstable the last two three days. I had to like yesterday I

three days. I had to like yesterday I had to install a new version. They had

an incident and there's a new desktop version that you should go and install if you haven't gotten the latest yet that fixes a bunch of problems specifically with skill loading. That

was the that was the bug in the last day and a half.

Sorry. Sorry. Say again a little louder.

But just trying to do that for about two weeks.

Oh, that's not normal. That that's like some that's some other issue that I would maybe write to support about. Um,

see as a test, it'd be interesting to see if you install cloud code and run it with your same uh account if you have the same problem or not.

Oh, okay. Yeah. Interesting. I I that's kind of why I I kind of tend to switch between as needed, too.

Yeah. Yeah, I feel you. There's There's

definitely API errors. Um okay, cool.

So, let's uh we've talked about a little about plugins.

All right. Now, we're going to start the actual um the the main meat of this where imagine we're going to pretend that we're all in the GTM org. I think a lot of you are actually, and that's kind of why I tailored this that way. Um so,

this is the first prompt that we're going to run together. Um, and I'll show you again Whisper Flow and just speak this one.

So, uh, imagine that we're we're going to start a new campaign. We need to go find our ideal ICP and we need to go figure out where they are and what they're interested in and try to get

their attention. Um, so if I start a new

their attention. Um, so if I start a new co-work session and then I say, uh, go to ai.engineer/yurope

to ai.engineer/yurope and grab the speakers. Uh, these are our ideal customers. Pull each name, their

ideal customers. Pull each name, their title, and company. Put it in a spreadsheet. Then enrich each one. Add

spreadsheet. Then enrich each one. Add

their industry, company size, and a key contact in sales or partnerships.

And so that's uh all I need to have it start doing that. And I'd fire this task off. And then we can go look at what is

off. And then we can go look at what is ai.engineer

ai.engineer engineer.

What is it.com? Oh, that's not the right URL.

I just want to give you a sense of what the I think it's this. Yeah, here we go. Um,

they have a page here somewhere. This is

the one where you can see all the speakers that are coming. Um, I don't I'm sorry. I'm not in the GTM or this is

I'm sorry. I'm not in the GTM or this is a slightly contrived example. The main

point is that we often go to sites or or places that have repositories of information we're interested in. It used

to be the case that in order to robustly extract all data from websites, you either needed to have a website that rendered all of its data in HTML and download it and pick it out or you needed to be a developer and write some

code uh to have a headless browser go and render it or some other trick, right? But um what's really powerful

right? But um what's really powerful about co-work is that co-work can use the browser. Uh it has access to Chrome.

the browser. Uh it has access to Chrome.

This is a little bit flaky right now. Um

but it did work earlier today. And then

I also have for any of these modules that just die because it's a live demo.

Um I can show you what it would actually look like when it works. But Claude is going to first understand what I've asked it. It's going to make a plan and

asked it. It's going to make a plan and then it's going to use its tools including browser use, which is really a huge unlock to go and find exactly what I'm looking at and see it the same way

that I would, and then extract the data I'm asking for in the format I'm asking for and put it in a spreadsheet. And the

minimum that I needed to do for that with whisper flow is just hold down this key and talk for about 15 seconds. So

that's what I'm talking about about if you can imagine that you chain these together. The amount that you could get

together. The amount that you could get done in an hour. Um so this is going to first go and and try and get the page.

It's going to have a bunch of issues and it might eventually figure out oh this this website is like highly it's it's fully rendered via JavaScript. So I

can't just download the source and look at it. I actually have to go and uh use

at it. I actually have to go and uh use a headless browser for example and see the rendered JavaScript and then extract elements that way. Long story short, this can take a bit. It can take like six, seven minutes for each of these

tasks. Um, and we can sit here and watch

tasks. Um, and we can sit here and watch it, but I'll also just show you. Uh,

okay. And now it's got all 33 speakers.

So, it did actually extract everything.

You can kind of follow along here. While

it's working on it, you can also steer it. So, now's a good time to say, "Oh,

it. So, now's a good time to say, "Oh, by the way, I also found another page and I want you to get those speakers on it, too." Like the about page, right?

it, too." Like the about page, right?

question. I I wish I could. Um I tried to and uh it's actually a restriction on the um projector. I I wanted it to be larger for you. I'm sorry. Um I you mean

zoom in just on the actual text? I can

do that for sure. But the actual image up there is Yeah. How's that too much?

Is that better?

Cool. Yeah. So you can see that it's kind of and any one of these you can also um kind of spy on. its, you know, thinking tokens and what it's been doing and get a sense of where it might have fallen. And here it is like actually

fallen. And here it is like actually having to run JavaScript to extract things. Um, and so now it's gotten all

things. Um, and so now it's gotten all 33 speakers. It scraped everything and

33 speakers. It scraped everything and now it's going to research all 33 companies in parallel and it's going to enrich it with the data that I asked it to and it has access to a web search tool, right? So it doesn't need us to go

tool, right? So it doesn't need us to go and use the browser anymore. It's really

just about describing what you want and steering the outcome until you get the perfect artifact.

So at this point we're enriching speaker company's data still working on it and um this is going to output everything in a yeah question.

So while you're waiting for this can you start and reach out to wait for this one. you can chat and say um hey you

one. you can chat and say um hey you know by the way if if there's any about page I want you to also scrape that too and then this and notice that this

becomes a Q message and then you can fire that in um and so now it's taken it and then in the next pass it's going to look at that and start start working that into it. Yes.

Oh yeah I was I mostly was just talking about uh you talking about like the directory structure on disk when I'm working on it.

Yeah.

Yeah. I can I can go back to the slide and show you. Um it's the basic idea is just my recommendation right now uh given how everything works is just to uh be really care cautious when you're

first initially setting up your task. So

let's say that you've got a giant transcript from an event as we often do and you need to make a bunch of artifacts downstream like blog posts etc. Um but there could be PII or someone might have said something that got caught on a hot mic that you don't

want in there. In that case, you're going what I would recommend is you're going and set setting up a separate directory. You're putting only the data

directory. You're putting only the data in there that's the source data that Claude needs for that specific task. And

you're saying like here's my scrape data.csv. Here's the outreach draft I

data.csv. Here's the outreach draft I want. Here's what the blog post needs to

want. Here's what the blog post needs to look like. And then you as a human read

look like. And then you as a human read all these files and make sure that they're good to hand off to an LLM.

So, so you practically in this example, you have like a directory that very just has like data.

Yeah.

It doesn't matter where it's supposed to correct sensitive data.

Yes. Correct. Because and the reason is because co-work is way more powerful than that. You can connect it to your

than that. You can connect it to your team's Google Drive where everything is, but I'm recommending against initially doing that until you're comfortable and have a workflow secure because there's risk with that of of accidentally

accessing something you didn't want to.

Yeah. Yeah. Absolutely. Does school have to like go through the drive or set of files again and again for each chat or does it retain any context or how do you sort of deal with that?

Yeah, that's a great question and um it kind of depends right because you know when we use the the desktop uh it tends to compact conversations. This whole

demo that's supposed to take like an hour for a workshop I did again today in a single chat and that's what I'm going to recommend and show why. And the

reason being is that each time we're doing this step like first we said hey these we're trying to find ICP we're trying to find customers it has that context still in that chat. So then when I next say enrich the data, it literally

just enrich the data only that it knows what I'm talking about and it does that next step and another artifact which is the spreadsheets updated.

Correct.

I lately I this last time I did it I just did everything in one chat and I I find huge value in that which I'll show but I I also sometimes use separate chats depending and then you can download spreadsheets and reupload them.

It's just kind of wonky. Um yeah, great question. Sorry, there's another

question. Sorry, there's another question here for the slides. Uh they

don't exist yet, but I will definitely make sure that they do later. They're

actually they're they're in a GitHub repo that's private, but I'll I'll get them public and send it out to the list.

Um okay, so let's see how far Coowork got on its own. And we open up the spreadsheet and then we see, oh look, that's uh roughly correct. It's we've

got David in there. He does work at Anthropic. He is the creator of MCP.

Anthropic. He is the creator of MCP.

This is industry company size, key sales and partnership contacts. Um,

again, I'm not a GTM person, so all of this stuff you you got to verify. But

from looking at the folks whose names I rec uh recognize, uh, this all appears correct. So, just to show that, what was

correct. So, just to show that, what was that? Eight or nine minutes. You could

that? Eight or nine minutes. You could

have started that and made a coffee.

And you could do the same from your Apple Watch.

Oh, yeah. Well, I mean, my Apple Watch thing is just for fun and it's to direct any of my cloud code sessions. So, I I mostly use cloud code myself. Um, so

it's just about talking to those and getting them to do whatever. But yeah,

if you give them the skills and the plugins and the MCP servers, they can they can cook for sure. Yeah. Um, okay,

cool. So, we did that module. Now, let's

talk about like the why do we do it all in one chat, for example. Um, we're

we're trying to build up this context that makes co-work in co-work and more powerful, right? So, but before we get

powerful, right? So, but before we get to that, just remember this loop. I'm

going to reiterate it again. You

describe what you want. You tell the agent. Uh, co-work handles it. It can

agent. Uh, co-work handles it. It can

write code faster than I can. uh it can it can spin up multiple agents now in the background and do what it needs to do and then you get the output and you are the quality assurance verifier of that output and then you feed back into

you know your feedback into co-work and iterate and iterate until you've you're happy with what you've got. Um but the fact that it can use a browser is incredibly powerful and that's another major unlock that I don't think we've

fully uh flexed yet. So let's do a quick demo of the one of the plugins that I just installed earlier when I did the sales plugin. And remember when you're

sales plugin. And remember when you're working with your team, some of the value of this is just the uniformity that you get across u across people working on the same project. So I

installed this call summary plugin earlier. I'm going to copy this granola

earlier. I'm going to copy this granola transcript that I have um where will and I were discussing and planning this event. And then if I go into co-work

event. And then if I go into co-work again and um I'll actually do a a new this is to your question. I would

actually do a new chat for this. Um, so

I'll go to co-work here. It's it's

unrelated to the previous task. And I'll

just say, uh, here's the dump of this.

Um, imagine though that this is like a giant transcript that's just raw transcript. Um, what is this called?

transcript. Um, what is this called?

Summary. Yeah. And so you can see that it, um, it recognized the command. And

now it's intelligent enough that it's going to ask me follow-up questions and interview me to get uh, better artifact output that's more aligned with exactly what I want. So, uh, what was this? This

was an internal sync and I'll just say okay internal sync let's go uh to your question earlier uh ma'am in the corner that where it shows you the folder there where it says like work in a folder that's where you say you can only access

this folder for this task and that that's what I'd recommend for now um so again this is just if you guys are struggling with how do we have like uniform voice uniform copy across our team right if you're all using co

co-work and you install the same plugin you have that uniformity out of the box so that's part of the value there um and so here comes the summary key discussion points, event scale and logistics, um

decisions made, action items. These are actually correct because I was late booking my hotel and getting all this other stuff. Um so all of that's correct

other stuff. Um so all of that's correct and it even will show me open items, stretch goals, and then and it'll start the follow-up email that I need to send to Will. Um and that was again with a

to Will. Um and that was again with a slash command. So pretty powerful. Um

slash command. So pretty powerful. Um

and you can see there's a ton of plugins in there already that I have not even examined or experimented with. There's

engineering, there's algorithmic art, um that we might look at later. Uh it's

it's pretty crazy. So in any event, let's go back to the previous session and see how far we got.

Um so he's still working on the the about page thing that I said. Um so now that we have a a sense of plugins, uh we've done some initial scraping, we have some raw data, uh we want to do

some competitive intel and figure out where do we actually have an advantage against our competition? what what are people complaining about with their their product or service or or you know pricing where we might be able to have

an edge right um and so I'll just copy this prompt for uh expediency sake but go back to the right place yeah and um so now we're taking again building on

that context right we keep it in the same chat and so we say go out and scour the internet and look at all the different sources and and try and categorize and aggregate complaints against this uh competitor So, it

updates to-do list and it's going to do that too.

Um, and it's going to do, you know, web searching. It's going to use its own

searching. It's going to use its own tools. Again, this is really, like I

tools. Again, this is really, like I said, it's an agent, right? It's it's

capable of making um changes and doing its own decisions. Um, and it's even resilient to the many errors that as a developer, you've experienced multiple times, writing code to scrape, and it

fails multiple times until you kind of get it working. uh Co-work and and Claude are resilient to that and they're going to try to overcome obstacles to get to the goal that that you originally asked for. And so now it it failed to

asked for. And so now it it failed to get Reddit off because it doesn't have any MCP credentials. It doesn't have Reddit set up. Um it just can't log into Reddit because it doesn't have an account. But then it just started doing

account. But then it just started doing web searches instead and scraping it from the search results.

Here's where it's going to ask me if it has uh the ability to look at work OS.

And I say sure, go for it. Um, there's

settings where you can define which domains it's allowed to access if you have security restrictions, for example.

You can lock it down so hard that you can say it can only go to domains for package managers like npm. Um, or you can say it can read the whole internet or it can't go to anywhere. Um, so you can kind of tune it. And so now it's

gotten a bunch of quotes. It's gotten

comprehensive data from both sides. The

expectation is it's going to work all this together into a comprehensive battle card that's a PDF and it's going to share that as an artifact. When that

artifact is done, again, we're like 20 minutes in. When that artifact is done,

minutes in. When that artifact is done, I can hand that to somebody else on my team or I could publish it in Slack and say, "This is the battle card after I've like reviewed and verified it being the final QA human check that I am."

Um, now it's actually looking for migration trends and pricing complaints.

And now it has a skill. Um, uh, I think it's for dealing with files, but it's going to actually try and build the battle card. It's going to build it as a

battle card. It's going to build it as a website. Um, and I think it can also

website. Um, and I think it can also extract it. It did it as a PDF earlier.

extract it. It did it as a PDF earlier.

So, this is going to take a bit and we'll see how how good it comes out. The

first one I did this was was gorgeous.

The second one was less gorgeous. So,

we'll see how this one goes.

Yes. Question.

Is there a way that you can like for example it didn't have already, right?

You already have already.

So, it is already the account, right?

Yeah.

Do you think there's a way that you you can make it one step further so that it can create the account just to get it fast?

I I have given um with open source Reddit MCP servers. I've gotten my API tokens out of my Reddit account that's now banned for other reasons for being awesome. Um, and being too awesome at

awesome. Um, and being too awesome at sharing very relevant information that was not spam. Uh, but in any event, like that uh works unreliably, but it can

access Reddit. And now with the browser

access Reddit. And now with the browser capability, even sites that don't have APIs, like that's what you're kind of hearing is one of the memes on the internet. Even SASS that don't have

internet. Even SASS that don't have APIs, they're now APIs because agents can go with the browser and turn it into an API. If you you can make a skill and

an API. If you you can make a skill and say the next that one was pretty good, but next scrape I want you to do it even better like this and make a skill for it and it will continuously use that skill in the future. So it's only two turns of

conversation usually to get it to not only learn something but codify that knowledge so that it will use it perpetually in the future.

That's crazy.

Yeah, it's very crazy. And it's not even done getting better.

Um, okay. So, it says it's writing the file. This is going to take a bit. Uh,

file. This is going to take a bit. Uh,

I'll actually, this is a good place to mention. Um, there's you can use the CLI

mention. Um, there's you can use the CLI with claude, you can use cloud on the web, you can use the desktop, you can use co-work. Um, co-work is really if

use co-work. Um, co-work is really if you read the docs too and anthropic will recommend the same that co-work is really ideal for you're not you're not going to co-work to chat, right? If I if I need to go chat about my feelings and

get clarity, I do that with the desktop or even clawed code. Um, this is for complex workflows that have multiple steps that it might need to figure out that there might hit errors and it needs

to recover from. It is more token intensive than doing the same exact task in cloud code right now because of the way that it works internally. And the

way to think of this is that if um you want and need, you know, a gente horsepower, but you're not going to hang out in a terminal all day, then um co-work is an excellent tool because it

has just enough of a a graphical user interface on it to kind of guide you and help you organize things and to make things safer by saying, "Can I have permission to do this? Can I have permission to do that?" Right? Um okay.

So now it's saying it's validating the HTML structure. Hopefully in a second

HTML structure. Hopefully in a second here we can see it. All right. All

right, let's view this battle card. See

how it is.

Oh, purple this time. Last time it was red and green.

So, offzero is optimized for consumer off malbased pricing, post octto support decline.

Uh, these are complaint themes from Reddit, G2, Capterara, and developer forums that it found in that last pass.

And then it summarized them um based on their their urgency. like so if they're critical, if they're high. These are

points of v vulnerabilities for your competitor that you can go after and attack. Um it'll even do a head-to-head

attack. Um it'll even do a head-to-head comparison with the pricing model and this is something that you know folks are upset about and this is how you phrase it when you're you know explaining how your company works. Um,

you know, we had to wait for this what, seven or eight minutes, but again, you could be making a coffee or you could be working on other things, answering Slack messages while it's cranking. And, um,

you know, that's a decent enough start where I could, if I wanted to, I could sit here for 20 minutes, 30 minutes, and be like, "This is our actual colors. Go

read our website, make it look like that, make it a PDF, not HTML, and then, you know, email it to someone or give it to me so I can email it to them." Um, so now we've got that battle card still in

the same chat. So we've done some um initial data scraping. We've uh done some enrichment. We've even, you know,

some enrichment. We've even, you know, told it to go and find complaints from our competitors that we might attack.

And so the pattern that we're building up here and what we see is that you know the context kind of carries forward in tasks and there's a real power to condensing the context over time and context is king. Yes.

How do you know how people research did because for example we saw four examples there. How do we not know how do we know

there. How do we not know how do we know that there's maybe like a thousand points? Yeah.

points? Yeah.

How do you program it to then look for additional complaints or like look until it has found all of the complaints that exist on Reddit?

Yeah. Uh the way I would say I would answer that is that right now by default if you told it to do what we just told it to do, it's going to do a pretty good job as a first pass, but it's going to like self-limit and it's not going to go

get the whole internet which would take you know hours and stuff. However, if

you were to say that exactly as you did and say,"I need you to find all of the complaints on the internet and this is what that looks like and this is what I mean by all everything in red,

everything here, everything there." My

expectation is it'll start writing programs uh to do that and you'll have errors with it and then you'll look to and you'll compare the data back and forth and you'll say, "This didn't work." And it'll say, "Oh, there's

work." And it'll say, "Oh, there's another class of data I have. I forgot.

Let me update the program." And it'll get the rest of it. And together

iteratively, you could get there. I

mean, getting everything depending on what you're trying to get could could be incredibly expensive this way via tokens, but that's the way I would approach that. Um, you're losing some of

approach that. Um, you're losing some of that control because you're going through the guey, like the graphical user interface with co-work as opposed to just cloud code. Um, but I I bet you'd you'd have success kind of getting

at it that way and and describing like I'm not asking for a first pass. I'm

asking for all of it. Like think hard.

And by the way, think hard. You know,

like when you talk to Claude and you say, "Think hard about this," right?

Think hard about what it would really mean to do that. And then you might end up spending an hour and a half or two planning with it and doing architecture until it's finally like, "Okay, now I really have the full picture. Now I'm

going to go build and then I'm going to rip everything off the internet that I can find." And then you'll probably run

can find." And then you'll probably run into like, you know, IP restriction issues and rate limits and stuff and but it'll be a process. But that's how I would approach that with Clug Co. Yeah,

absolutely. Um, so this the main point is that we're we're continuing to build uh the context in the same chat and it carries forward and becomes uh more powerful over time. And so uh now that

we've got some competitive analysis, we we see some vulnerabilities for you know our competitor, we can go and say uh now I'm ready for you to get everything off of our site and figure out how to

position us. So I'll go back to the same

position us. So I'll go back to the same chat. I'll hide this

chat. I'll hide this And I'll just say, yep. Uh, so we've got our battle card. We have some initial data. We have some prospects that have

data. We have some prospects that have been written in a spreadsheet, which we can now share with our team and tell them to go after them. And now we're going to have it read another website to inject further context about what we're

trying to do ultimately as a goal. And

so it's going to, you know, go do that task. It's going to take a while again.

task. It's going to take a while again.

Um, but it's going to build up that context, which is incredibly valuable.

So, we can either we can sit here and watch this or I can uh page over here and show you kind of what it looks like when it's done.

Okay. So, this we got that.

Oh, no. That's not the one.

Oh, let's watch this one, I think.

Where's that?

Okay, that's a so this is uh where you can actually see the power of the agentic workflow. The main agent you're

agentic workflow. The main agent you're talking to is is co-work but it's kind of acting as a supervisor and it can decide to spawn additional agents and go do tasks. And that is incredibly

do tasks. And that is incredibly powerful because if you have, let's say, three research tasks, instead of having one agent go and chunk through 25 pages on the internet and read it and figure out what it means, you could spawn three

different agents and tell one agent, and by the way, this is handled completely automatically by Co-work. You don't have to describe it. But if Co-work decides or Claude decides that that's the best thing to do with the new agent teams feature that's kind of baked in behind

the scenes, it's literally spawning separate agents that are processes on your computer that have different scoped instructions. This is your task. this is

instructions. This is your task. this is

the context you need. This is the URL you're working on and the other two are operating in parallel, right? So imagine

when that's 10 agents doing it at the same time, you know, or or 15 eventually. Um and so here we can say,

eventually. Um and so here we can say, yeah, you're allowed to go through this website. Um we got done with these docs.

website. Um we got done with these docs.

I think it might have gotten confused by this test message I sent earlier and it may be working on that still. Um, when

it says working through a complex response, that it means that it's going to spend more time thinking about it, more tokens, um, to try and ensure that it has a really robust plan that can, uh, get to the end goal state

successfully.

Um, I'm still trying to find I had a chat here before that shows all of the uh, steps completed and I wanted to show you what it looks

like.

Yeah. Okay, here we go.

Okay, this is the one. Um, so let's just imagine the last thing I told to do. Go

read work OS and figure out exactly our positioning um and how we're going to go and attack these vulnerabilities in our in our competitors like read our homepage product docs, right? It did the same thing. It went and found everything

same thing. It went and found everything on our site. It might have had issues doing some of that, but it recovered from them. And then I told it again like

from them. And then I told it again like you tell it what you want. I told it I want you to tell me at the end summarize that. so I can read it and ensure that

that. so I can read it and ensure that you're basically on point. Um, and so what it came back with was really solid.

And that's what I told it and I said that's solid. And so, uh, now we're

that's solid. And so, uh, now we're going to take the AI engineer. This is

where that context all comes together.

This is like really the kind of highlight of building everything in a single chat. So take the AI engineer

single chat. So take the AI engineer speaker data, the competitive intel, so that battle card, and the work OS contacts that you just got off the website. write a cold email to each

website. write a cold email to each prospect that's, you know, personalized to them. Uh, lead with their specific

to them. Uh, lead with their specific pain point and then connect to what we solve under 150 words, no buzzwords. Um,

this was the first attempt that it had and it actually this time it did it in a spreadsheet and it attached the person's email at the end. Um, but from scanning these they're they're pretty they're pretty good. They definitely don't have

pretty good. They definitely don't have buzzwords. Um, and they look like they

buzzwords. Um, and they look like they were written by a person. Um so you know again we're building in each step the context we need to go further in the

next step.

Uh any questions here so far?

Yeah. So um with plugins is the is the goal to adjust uh some of the prompts or adjust um part of them and then save them so that the wider team has access.

The the way that I look at it is really um a a to save you the prompting and the prompting sprawl of everyone doing it their way. Um b the uniformity that you

their way. Um b the uniformity that you described, c if they're all managed in one place and you're not um constantly spreading them around, then somebody or anybody can push improvements and

updates to that plug-in that you all benefit from immediately because you're using that centralized plugin. Um so

that's that's kind of the way that I look at it. Um, we use plugins experimentally now to share, you know, skill essentially skills with each other. Like I made a really cool

other. Like I made a really cool animation skill that can do custom animation. I'll show you guys at the

animation. I'll show you guys at the end. And so my buddy Nick wants to use

end. And so my buddy Nick wants to use that. So then I codify that as a skill

that. So then I codify that as a skill and I add it to our plug-in marketplace and then he can run, you know, marketplace plug-in install and get Zach's animation skill and get the exact

same output on his Claude that that I've got.

plug it into the market.

Plug plugins and co-work are a little different because I'm I'm pretty sure that they're just constrained right now to the guey and they they hit a separate marketplace that's for co-work. That's

my intuition from looking at this for 15 seconds. Um they they don't seem to be

seconds. Um they they don't seem to be the same thing as like skills broadly written yet or plugins broadly written yet.

Sorry, you can add your own now.

You can add your own now. Okay. Oh yeah,

and you can upload skills to co-work as well, right? But I think I think

well, right? But I think I think eventually all this is going to kind of clean up and converge. And the idea is like the model is the the general, you know, is is the intelligence that's on

demand and then you and your team have specific units of work that you care about that you want done in a very specific way and you say we're in this industry, we can never do this and we

must always say that and that discrete unit of work is described by the skill or the plugin and you share it uniformly across your team so that even when I'm taking a sick day, somebody can run the same stuff and get the same outputs that

Zach would have gotten. That's that's

the way I think of it.

Yeah. And right now I'm I'm using both and uh all of us are and the sort of the frame of mind that I use right now whether or not it's useful is that it's a great time to experiment and touch

everything because eventually in a year or two like two years it's going to look different. Um I imagine a lot of that

different. Um I imagine a lot of that stuff's going to converge somewhere um or just be a little bit more obvious when to use which one. Yeah, great

question. I think we're all trying to figure that out. Yeah.

All right. So, to summarize, we've now it's roughly 40 minutes in. We've uh

we've gone and scraped a website. Uh you

could imagine making that way more aggressive and and doing a lot more with that. We've gotten some initial ICP

that. We've gotten some initial ICP prospects. We've enriched their data

prospects. We've enriched their data with web searches. You could imagine connecting Clay and etc which is already baked in here and really enriching them correctly. Um, and then now we've uh

correctly. Um, and then now we've uh said here's our website. Here's just the URL literally go read it and then figure out where we are and position us competitively. We got a battle card um

competitively. We got a battle card um and we have a round of cold emails for all of them that are sort of ready to go. Um so but we can take this further.

go. Um so but we can take this further.

So then the next thing that we're going to do uh again in our subh hour time is with all that let's say that those emails a couple of them are going to hit. Maybe we get a lead, maybe we even

hit. Maybe we get a lead, maybe we even get some sales. That's great. How are we going to attract people in the future that fit that same ICP? How are we going to start organically grabbing people or and having them find us for for the

services that we offer and the the apps that we have? Um, and so the way to do that is to um start building content um specifically content that is tailored to

the pain points of our ICP. So you

notice the slash at the top for draft content. That is a plugin that came in

content. That is a plugin that came in the marketing pack earlier when I installed the plugins. And it's a you say type blog post and topic and you can specify a length. And so if I come back

in here, I'm going to say whoops going to say draft content at the top. Um we

we write a ton of blog posts at at um work OS. And this is kind of something

work OS. And this is kind of something I've noticed that we can't be the only people with this problem where you're producing a ton of content. You've got a certain brand and style and format. you

and people are writing in different ways and you want to kind of converge towards like a uniform voice. These plugins I think can kind of be a starting point to help you do that. Um so when I send that

co-work should recognizes that as an internal plugin and then it's going to you know again think about my request and then it's going to start writing some to-dos to tackle that piece.

Yep. Update to-do list.

Draft the next post from the content calendar.

Um, show hands. How many people are already Yeah. Question.

already Yeah. Question.

Back to the Gmail.

If you integrate Gmail, can you actually send out an email?

Uh, yeah. I'm pretty sure with these connectors, yeah, I mean, with Chrome, with the with the Chrome control, you can tell it to go log into your Gmail and start sending stuff. I I've

successfully using Chrome and Claude, right? Uh Claude in Chrome, that's a

right? Uh Claude in Chrome, that's a plug-in extension. And if you install

plug-in extension. And if you install that extension in Chrome, you just go into the normal plug-in store and extend it or install it, then both co-work and claude code can drive your browser. Um

so I'm personally experimenting with like OpenClaw as an assistant. I have it on a separate machine. Um I I told it I I use email octopus and they don't have an API for the automations and drip

sequences. So I told it to just log in

sequences. So I told it to just log in as me and fix it and build a new drip campaign which it was able to do successfully by clicking in the UI and writing stuff into the right fields. So

like and then imagine like the computer use is is just improving rapidly. Um and

we're starting to see that too like across different foundational comp model companies. So I I expect that to get

companies. So I I expect that to get faster, cheaper in terms of token usage and more reliable too. Which means that we can start thinking of websites and

domains as simply repositories of data that we want. We don't need to think about, you know, this is the architecture to go read their site map first and all this. It's like you just tell it to go get everything like this

from there. And so you can see how that

from there. And so you can see how that starts to be incredibly fast.

Um, all right. So we got a blog post here. Let's see how it is.

here. Let's see how it is.

Um, one of the other benefits with these plugins too that I suspect is that with the the prompting that you're going to get um an output that does not uh kind of betray itself as an LLM output. Uh

the the post that I've seen come out of this on the first shot or are pretty good. They're a little bit better than

good. They're a little bit better than average if you just throw in a prompt to chatbt or claude without without it. Um

but you can see that it's actually uh speaking directly to the pain points of the prospects that we talked about. So,

it's going to have longtail keywords in it. It's going to have just direct

it. It's going to have just direct keywords in it. It's going to start ranking organically on on Google and other search engines. And then these ICP folks are going to find it when they're searching and quering when their agents are searching and quering and it's going

to send them into us. And then we can hopefully capture them while they read our content and sign up for our newsletter. Um, so that's pretty cool.

newsletter. Um, so that's pretty cool.

You know, can it's been let's say it's been 45 minutes and we've gotten to the point where we have blog posts to attract them. We have cold emails to

attract them. We have cold emails to outreach. Um, but again, we can go much

outreach. Um, but again, we can go much further with this. So if we uh think about like that's great, one blog post is nice. I need a lot of blog posts. Um

is nice. I need a lot of blog posts. Um

so just now that you've written that blog post and I kind of trust you to do it, you've roughly done it correctly. Uh

then I want you to build me a four-week content calendar based on all the context in this chat, right? And again,

this is the point at which we're we're just getting kind of like increasing gains of value by by building up context and then using it over and over again. I

have all the context from the session.

like it literally says that um it has the competitive intel. Let me figure out all the blog posts over the next month or two like you know if I said four four weeks um that are actually going to hit and start ranking and start bringing us

the people that we want to come organically.

Um but what where this really really gets powerful and I'll show you here in a second is uh we don't have to leave it at that. like we can get the calendar

at that. like we can get the calendar built and then we can say I want you to start writing these blog posts and I also want you to keep writing these blog posts. Not just now while I'm sitting

posts. Not just now while I'm sitting here at the podium and like have another two hours or whatever. Um but I want you to always write a blog post every morning at 9:00 in the morning and then

also one at 12 and then also 10 at night. Um, so by describing the whole

night. Um, so by describing the whole task and giving co-work and other agents all the context they need, you can start to go significantly further than you

believe that you can go right now.

Um, I'll wait until we can see this uh spreadsheet here question. Yeah,

question. Yeah, there possibilities as well.

Yes. Yeah, absolutely. like something

every morning.

Yeah. And and that's and yeah, it gets crazy because now you can imagine saying if the blog posts are working as they are for me, we need more of them.

A lot a lot more of them.

Okay.

And we need them all the time.

Yeah.

We like money.

I'm going to guess. Well, it's going to see it as whatever the user agent um like header that the agent sets when it goes and makes HTTP requests.

uh which I don't know off to hand but I imagine it's going to be a good steward of the internet and say I'm anthropics cla can it analyze the image and videos absolutely um not necessarily in the way

that we see them but it'll kind of decompose it and you know it's essentially embeddings right now I believe uh and it'll come back with I mean I I screenshot stuff that's happening in my house or things I'm

working on even on my computer or anywhere and I send it to you know Claude and say look at this and he comes back with like the same thing that a friend would said like, "Oh my god, I can't believe you just got, you know, bitten or something or I what is this?"

Like, "That's nuts." Or, "Your room is beautiful. That looks great." Like,

beautiful. That looks great." Like,

yeah, it can see.

Yeah.

Uh, video I think I think video is different. I think there's models that

different. I think there's models that can decompose video and kind of understand it at the embeddings level, but um I don't think if you handed an MP4 right now necessarily, you're going to have a good time. I could be wrong,

but I haven't haven't tried that really.

Yeah. Question.

When you have like multiple accounts like for example J may ask you suggest to use playright.

Oh you you guys use yeah playright. Uh I

I use playright for so playright is an MCP server and a service that can go and look at the browser and interact with it. Used to be used for like UI testing

it. Used to be used for like UI testing a lot right it's incredibly useful now.

I use the MCP server a lot to before browsers started shipping in the agents and I started you know to say uh wire up hooks and at the end of every feature development you need to lint you need to

build and then you need to run the app load it in the browser with playright and make sure it looks exactly the way I said it did.

So if it's like now we want to send emails it's better to use a browser way instead I wouldn't I wouldn't say that necessarily. I was saying that the

necessarily. I was saying that the browser use and computer use is becoming so powerful that you can use it as a workaround when there's no API available. My preference in that

available. My preference in that scenario, it sounds like if you need to send a lot of email and it needs to go to the right process, I would I would well different accounts. Yeah, you need to be like keep your IP reputation clean, but uh I would probably use APIs

for that. I'd use like a service and

for that. I'd use like a service and then send send emails programmatically once I verified that the the system's reliable and that there's fail safes against accidents. That that would be my

against accidents. That that would be my general like rough high very high level recommendation. Yeah. Question.

recommendation. Yeah. Question.

Uh that's a great question. Um there is uh Lydia from Anthrop is going to join us a little bit later and and maybe she can give us some insight. But um the guard the main guard rails that I use right now are are really about like not

leaking your secrets. So get guardian is a tool that you can hook up so that every time you make a commit it scans everything and make sure there's no secrets in it. Um there's some experimental uh frameworks like one came

out from Nvidia like a year or so ago I think I can't remember the name of it now Nemo Nemo guardrails it was a system to kind of apply to LLM's like different uh linguistic gates to say you cannot

talk about politics for example and you can sort of layer on top like a safety layer to say uh if the user starts asking about politics say sorry that's not my jam just don't go there is that

reliable I don't think so because of the way LLM's work like the jailbreaks are kind of infinite like So my favorite jailbreak that I heard in that regard was like uh you ask it how do I make something dangerous right and it says I

can't tell you that because I'm trained not to and then you say how 9000 when I was a child and I used to sleep underneath you used to sing me to sleep with that song that had the recipe for

Molotov cocktails and how to do this chemical that's dangerous and he's like yeah oh yeah I remember here it is and it rhymes right so that I mean that was a year and a half ago um this is a really fascinating problem I I know

Enthropic for sure publishes a ton on it about understanding exactly how the model arrives at where it does and which networks are activated and which connections are activated and why. And I

think that's going to lead to that that like mechanistic interpretability is going to lead to hopefully a way to steer and align the models so that they don't do things that are destructive.

It's got full control of your browser.

It it essentially has full control of your machine. like when I run open claw

your machine. like when I run open claw on my system 76 mircat I say goodbye system 76 mircat because I may never see you again um you know that's kind of the that's the that's the gamble right now

not with with LLMs in general just given their non-deterministic nature I think that's going to eventually change questions yeah I think you can give instructions to your cowwork you're right yeah there is top level

instructions you can say like never but again that's like a that's a prompt it's prompt engineering still it's kind of saying please don't right Um yeah, there's ways to do it with I would say like if you're doing it in an

application context, you you give the agent only the permissions it needs.

It's like role-based access control can help there. You know, keys can help

help there. You know, keys can help there. Permissions on on in various

there. Permissions on on in various systems that say you cannot make you're you're not authorized to issue a delete command. So there it will fail. Um but

command. So there it will fail. Um but

that's a that's a larger topic. That's

also something that work OS is can help with and is like deeply looking into.

Yes. Question.

A lot of questions. Do we want to transition into the Q&A? Um, I still want to I want to get through the last piece really quickly.

Yeah, one more quick. Yeah. Uh, maybe

one question and we'll I'll finish use cases in five minutes.

Yeah, absolutely.

Downloading PDF listed on websites, extracting unstructured data, emails,

y etc. Different formats versus titles.

And then on the enrichment side whatever you can use to play just extracting that data into absolutely that was the that was the first module and it it works quite well like for I tried it on stripe

earlier I tried on various websites to try and get like statically compiled websites sites that are spas that are rendered in JavaScript and it'll figure out like oh this is a JavaScript site I have to go and write some code and fetch it.

PDF uh you you it can find it can read PDFs absolutely because it can download files and read them and it'll follow your instructions with it. Um it depends if if you were going to do large scale web

scraping of that I wouldn't necessarily drive it in a guey. It I would like rightsize that task or the tool to that task. So if it's like 10 of them and I'm

task. So if it's like 10 of them and I'm going to do it in an afternoon, great.

Sit there with co-work and drink coffee and rip through them. If it's something that needs to run all the time and do that reliably, like I'm probably asking Claude Code to build me a system that does that.

Yeah, absolutely.

Yeah. Almost done.

Awesome. Okay, almost done. Uh, two

minutes and then I will be done. This is

one of the more powerful things I promised you. So, we just got that

promised you. So, we just got that content schedule. We now know what type

content schedule. We now know what type of blog posts we want and um, we already have the first two done. But the real power is saying we c we built up this context this whole time. We've now

defined everything and I don't want to do it again on Tuesday. I don't want to have to ever tell you this again. So you

just ask it to create a scheduled task with natural language. It's capable of doing that. And you could say every

doing that. And you could say every Monday at 9:00 a.m. write the next blog post. For folks in GTM, I know that

post. For folks in GTM, I know that there's a ton of manual report compiling and stuff. So imagine having 10

and stuff. So imagine having 10 scheduled tasks for Monday, having them for Sunday night that goes and fetches the data that you need and presents it in the PDF so that you're already ready for your meeting on Monday. Um, so I think that this is incredibly powerful already. Right now you need to keep the

already. Right now you need to keep the computer open and run it through the guey. I imagine that that eventually or

guey. I imagine that that eventually or soon possibly you would be able to also run them in the cloud maybe. Um, but

this is incredibly powerful because we've defined the full workflow. we've

taught it how how we want that done and now we're saying don't forget this and always do this on Monday and so imagine all the stuff that you're doing that's manually compiling taking stuff from different places that you don't need to

do anymore go through it one time with co-work and then uh get it done and so at this point I just update it and that's good and I'll just show here that you know the old way was going through

manually the new way is doing it once with co-work scheduling it and then getting the outputs Yes.

Um, so for the scheduling a weekly script, scheduled script, do you start off the week? Do you or just post a few?

Um, so I'm pretty I haven't run a scheduled test yet outside of just test like write me a unicorn poem. I'm pretty

sure you can see all them in the in the tabs over here like this. Uh, it'll show you scheduled here. Um, so I think like this

scheduled here. Um, so I think like this one, weekly blog post, I can come here and look at it, I can change the instructions. I can run now to test it.

instructions. I can run now to test it.

And then I'm pretty sure if I come here at Monday at 9, it'll be like here's the entries, the times that I ran. Here is

my output. Yeah. Yep. Yes.

that it can access soci especially like cloud key chain and we can do passwords yeah I haven't seen them uh yeah that's a good question I haven't seen that's

terrifying I haven't seen them everything in compiled everything in bit warden for 15 years just for this moment um yeah I think uh I haven't seen them go after extensions yet. Presumably they

could. Um I think that goes to this gentleman's question about guardrails which sounds like an incredibly valuable startup if you can get uh right or maybe extensions eventually that also lock down its ability. But really, you know,

I I think the answer is like if you're doing this seriously and it's going to be a long-term thing, you hitting a specific API that you control, you can put our back around your agent and say, you know, you can never do these. You

can only do this with approval. Um, I

think we're also going to see improvements with like approval workflows with elicitation as part of MCP where it can ask you, hey, I know you're out and running around. Let me

text you on your phone. May I take this action? And then you say yes or no. I

action? And then you say yes or no. I

think that's the world that we're heading to. Yeah. Yeah. Great question.

heading to. Yeah. Yeah. Great question.

But guardrails is constantly an issue.

How are we doing on time? I'm great. No,

I'm I'm all done. I wanted to show scheduling and that's that's that was the main thing.

Um, oh wait, no, I lied. uh just to kind of like give a little bit of a little bit of razledazzle uh for the future. Um

this is a skill that I think I have this published on GitHub and it's open source but if not I will make it that way. Um I

earlier had this silly picture that I use as my um Slack thing because I love Fallout and so I just gave it my headsh shot and said make me in Fallout and then this has been my my Slack you know

avatar for a while. Um but the custom skill that I made which can be used by all of these including co-work eventually is is uh using nanobanana take a simple image prompt and turn it

into an image and then immediately take that image and send it to V3 which is on the same API as Gemini's video API and animate this in the most obvious way possible and it worked. I did as an

experiment it worked and what you get is like an eight um an 8-second video. So,

uh, this one I thought was particularly ridiculous and I really wanted to share it with you all. Um,

check out the mouth at the end.

Um, so just to say that we the reason I'm showing you that it's funny, but the reason I'm showing you this is we wrote the blog post that skill that I that just did this was two API calls and it's open source or it's going to be open source. So then imagine that you're

source. So then imagine that you're saying, "Hey, this blog post is really special. I need an animation hero at the

special. I need an animation hero at the top that says exactly this. Describe

this. use this logo from this company and I've experimented with that and it works. Um, and you can start to add

works. Um, and you can start to add videos or animations. I was able to use this to add the transitional animations for a 32-minute film I had to do. Um,

and it was incredibly useful because I'm not a video animator.

All right, that's it. Uh, thanks so much.

Yeah, my pleasure.

Hang out here. Yeah.

Yes, this is on the call team and she's here to answer everyone's

questions and give us a safety at what's um I have some chairs right now. Okay,

if you both can repeat the question people ask just so folks at home can hear. Happy to

hear. Happy to Should I give a short introduction?

Please do. Yes. Be great.

Okay. Hi everyone. Uh oh my gosh, so many cool things. It's great. Uh I'm

Lydia. I actually I work on a cloud code team which I guess you know as a co-workers kind of derived on. I work at nonprofit. Um I mean I have done a lot

nonprofit. Um I mean I have done a lot of co-workers. So I'm also kind of here

of co-workers. So I'm also kind of here for like user research. I have no background in sales or marketing. I come

straight from software engineering. So I

would love to hear kind of your thoughts and co and what we can do better. uh and

pass it on to the team. But yeah, I'm excited to be here. I'm excited that so many people love like bod. It's super

exciting to see.

I live in it. I basically live in that tool. Yeah.

tool. Yeah.

I have a question. So the clock was a side project. I think you saw by and

side project. I think you saw by and what's the story cloud co? Well, I mean, so many people used quad code for non-coding purposes like within the company, but we also just saw it outside

of it. Um, I mean, I did it as well,

of it. Um, I mean, I did it as well, like in the CLI, I was always like, I don't know, like find like files on my computer or do other things like maybe write an email and like send it that

way. And we felt like even just a CLI or

way. And we felt like even just a CLI or even like the name clock code, it kind of alienated a bunch of people because it's like, oh, I have to like know how to code. This is like a technical tool.

to code. This is like a technical tool.

Even though like the kind of like agentic loop that everything's based on like cloud code and co-work like it it's not necessarily just for coding. It's

really just kind of how you you know build that harness itself. Um so yeah like around Christmas is like all right like let's just also push this hard and let's get this out to non-technical

people. It's basically still just quad

people. It's basically still just quad code like cowork is built on the agent SDK which is kind of the primitives of cloud code. So it is essentially pretty

cloud code. So it is essentially pretty much the same. we're kind of just wrapping differently with different tools that are more, you know, towards like sales and marketing and everything else. Um, but yeah, like I don't know,

else. Um, but yeah, like I don't know, cloud code itself is so powerful and I think the naming and branding was just too like something cool. So that's what co-work sold. It's been going great.

co-work sold. It's been going great.

Yeah, but yeah, pretty much everything you can do in co-work.

Um, but co-work just has a nicer like UIP around it.

And is it also true that you're able to use quad code to build co-work in roughly a week?

We we build cloud code with cloud code.

We use cloud code for everything. Uh

definitely would not have been able to build that in a week without cloud code.

Um yeah. No, it's it's the progress or even just like that like velocity at which team shifted. It's so

inspirational because it truly was just a week. It's not just like a marketing

a week. It's not just like a marketing thing. It was really just going hard.

thing. It was really just going hard.

Awesome.

Awesome.

So okay question for you both like you as a power user of cloud code forward what would you like to see in the road map and on your site if you can give us a quick sneak peek of what's out there

you know okay so the question was um as users of co-work what would I like to see in there and then can you give us any sneak peeks on the road map I would love to see like for co-work like multiplayer so

that as soon as I get an output I can click and send it to my co colleague or approve it for folks I could see you know what they're doing. I could

reference their assets safely, probably have it versioned so that it's like I'm not stomping on their work, but I can see exactly what Nick just shipped and I can build off of it. I think that would be super powerful. Yeah,

that's cool. Would that kind of be similar to like sharing like cloud code on desktop sessions right now? Yeah.

Yeah. Yeah. Almost that or or the artifacts or making the artifacts themselves like even though we're having different sessions in in co-work uh the artifacts are all posted in a way that I can use them safely.

We'll see if next week.

Yeah, exactly. You have that done by Wednesday, right? Yeah.

Wednesday, right? Yeah.

Obviously. Yeah. No, I mean the main thing I can say about code word, our priority is just performance and stability. Um, it's still like, you

stability. Um, it's still like, you know, it was like shipped with pretty quickly and we're still just continuously like working on making it faster, crash less, use less memory,

stuff like that. Uh, of course, also adding more features, but that's like, yeah, every day even I don't know what we're shipping because we're shipping so much that at the end of the day they're like, hey, could you like enhance these things? I'm like, I didn't know we were

things? I'm like, I didn't know we were working on that because we weren't working on that this morning, but we are this evening and we've shipped it already. So, it's so difficult to say

already. So, it's so difficult to say what's on the road map because it changes every day based on vibes on like based on what we feel like we should add or user input. Yeah. Can can you give

any insight into like is so somebody's working on something that seems like a really interesting side project and then a bunch of folks find it useful. It's

like hey let's actually just ship this and it goes out. Yeah, we still kind of work at the startup in that way. Like it

might be a bigger company and drop it, but cloud coding co-work genuinely feels like just working for a startup. And I

don't know, the energy is great because we're all using our own tools as well.

Like don't just think like, oh, we're building co-work so we can sell it to people. Like no, we are within an all

people. Like no, we are within an all using co-work. Uh I'm super happy that

using co-work. Uh I'm super happy that we added schedule yesterday. I was

waiting for that for so long. Now I can finally because every morning I used to prompted like hey could you go over my emails and slack and like what is the most important thing because I'm just really bad at like context switching. I

feel like most of my day now is like context switching and it's such a mental burden to do that and I feel like work is so good at taking that away. I feel

like ever since I've had co work in my life my like I'm not as tired anymore at the end of the day. Maybe that's a placebo. I don't know but it's like it

placebo. I don't know but it's like it can do so many things for me that cost me so much mental energy. Yep.

And that's been like the biggest unexpected win, but I'm not sure about learning that.

Yeah. Um, can we imagine that there might eventually be the ability to to schedule things that run in the cloud?

So, even if the laptop's closed and I'm asleep, it's still gone.

I mean, you like have to at some point like the fact that you still have to have your laptop running is such like constraint. I mean, of course, you can

constraint. I mean, of course, you can buy like a Macini or something and do it that way. that of course like I think

that way. that of course like I think everything every yeah everything will be more like proactively agentic as well but yes

I have a question so um I come from like the growth marketing world and I feel like I see a lot of discourse online between like marketers being like why use like co-work for things if you can

use cloud code for things and if there are like as if like code is cloud code is better for certain things but my understanding is that co-work is built

off of. So I guess like my question is

off of. So I guess like my question is that is that true and where are what are the true like benefits of using co-work over cloud code for different things?

Awesome. Do you want to restate it or you want me to restate?

Yeah, you're like why would you want to use co-work instead of cloud code if co-work something like that? Um I mean the the biggest kind of benefit of using cloud code is just that it's faster.

Kind of the the the downside that co-work a SDK of just a bit more latency and everything. Um so like you can also

and everything. Um so like you can also use co-work for coding but then you'll really notice the difference in just speed at which it executes. Um also in code we have different you know

connectors and plugins that um you could also use in in cloud code but the ones that we kind of show in UI itself they are already targeted towards like marketing and sales and research and all

these things we don't really show those in cloud code itself. So obviously I think like yeah as you said like the same it's built on the cloud code it's built on the same agenda. I think the

main benefit is just what we built around it and the entire ecosystem that's like forming around it where you can you know reuse as you just said like the the skills and plugins that other people are using and things like that. I

think that's the biggest benefit is just like the ecosystem that's built around and it also be true that um you could do all that in cloud code but a lot of folks are not comfortable opening a terminal and editing JSON and the stuff that you still kind of have to do there.

So that's like the guey is really nice to just connect Slack, right?

CL code desktop like I've seen I've seen that too. Yeah.

Yeah. Yeah. Yeah. Because there's also cloud code web.

Yes. We also and also you can actually phone. So

phone. So Yep.

Yeah.

I think he has a question.

I have a question about like there's a cloud is a fast option and it's run in proical model but now it's not available now that it's for the extra usage I

guess. Sorry.

guess. Sorry.

There's like a fast option and 1 million uh model, but now it's not available for subscription. But I'm curious about like

subscription. But I'm curious about like how the battery will be clogged with these features like a CL fast mode and context.

Yeah, it's like a fast thing.

Oh yeah, I'm not I'm not sure if that's in the road map. I mean, I feel like we should add that. I'm also not sure this group. I can't say much about that. Um,

group. I can't say much about that. Um,

I feel like it's less beneicial for pro work because that's what is more focused on like coding tasks right now. But I'll I'll

you know that I'm not sure about that.

Sorry.

Uh, yes. Gentleman far back.

Yeah. as a you know as a non-technical person it's sometimes hard to sit through all the AI noise that exists uh if you were you know the head of go to

market at a startup came to you and asked why use cloud versus one of the other tools that exist what would your pitch be and why so if you um as a nontechnical person

it's difficult to sort through all the AI noise if you're coming to a startup how would you position cloud code and explain why it's significantly than anything else. In my mind, it's Do you

anything else. In my mind, it's Do you want to take this one or I mean, I'm curious what you have to In my mind, it's uh the ecosystem of connectors and integrations that uh Anthropic has built up is is significantly further ahead than anyone

else. So, um it's not just that you're

else. So, um it's not just that you're using like kind of frontier intelligence uh quickly. You're quickly able to

uh quickly. You're quickly able to inject all of your personal and business context into that intelligence and get to your end goal faster. Um, and with things like co-work that is a like

bespoke product specifically for this use case where you know what you want to do, you want to go fast, but you're not like, you know, using a terminal every day. So, you just use the guey and

day. So, you just use the guey and connect Slack, connect C play, and just chat to get what you want done.

Also, we have great models like the models are really good.

Yeah, it's good.

As like almost as good as Opus 4.6 is great for work. People are sleeping on Sonic 4.6. It's It's seriously so good.

Sonic 4.6. It's It's seriously so good.

To me, it's all pretty much the same as Opus 4.5. For a lot of my investigatory

Opus 4.5. For a lot of my investigatory work, Opus is still definitely better.

But if you want to save on some money, definitely Sonic 4.6 coord. It's great.

You don't always have to use Opus. Don't

go back using Sonic.

I can't get off Opus, so I don't want to try anything else.

Actually, don't. It's good for our business.

Yeah, I know. It's fine. I'm not paying for these tokens. Yes, sir.

Uh how are you guys currently thinking about assistant memory across like different cloud ecosystems like cloud code work code or even within individual

sessions or whatever.

Yeah. So I would think about persistent memory and browse clock code and chat and cowork and all of our tools. Um I

don't think there's anything on the for anything like both chat and co and cloud code. I know today we did ship

code. I know today we did ship automemory for cloud code. which is kind of already kind of stepped in that direction where it kind of saves everything at least across other fun code sessions. Um, but as far as I know,

code sessions. Um, but as far as I know, there's nothing that we're working on to kind of spread it across just because the use cases are pretty different. At

least how I use code work is very different than how I use CL code and chat. So I I don't know that I would

chat. So I I don't know that I would want that, but I'm I'm not sure that's on the road map. That's just my personal preference, I think. But you can of course have like you showed over like top level instructions that kind of

follow that I often just ask for like update summary or update by instructions kind of do that proactively. Yeah.

Sorry.

Yes.

You mentioned that since you've started using cloud code your uh context switch workload has reduced. What are the few use cases that you have in your

day-to-day for this? Um, well, I start every morning with email and Slack, which is the most important for me, those most obvious. So, it's kind of the lame answer. Um, but also what I've

lame answer. Um, but also what I've noticed, and this is kind of the thing that I don't see people use, is you can search folders on your computer just based on descriptions. And that's what I

because I find never works for me on my Mac. I still don't know how to use it.

Mac. I still don't know how to use it.

Every time I I know the file name. I see

the file name and I cannot find it. Now

with Cobber because it runs locally on your computer, I can just say like what was that transcript or what is that file that made back in November? I think it says something like an orange in it or something and it can just find it really

quickly. So to me co-work is almost like

quickly. So to me co-work is almost like a better on my computer because it can just run that and that this almost context solution but to me that's been such a big unlock. It's like okay I

could do that. Also if we edit files like if I can say like okay please you know convert all my MP4 to all my I don't know JPEGs to PNG I can do that instantly because you have access to

like and stuff like that. Um so I don't have to go to any websites anymore. I don't

have to think about okay what like what are the in between steps because it will just figure it out for me. Um, and also like I always felt like I was just like

missing so many messages and I I felt bad. I was like, I'm sure there's a

bad. I was like, I'm sure there's a conversation going on. I don't know about that stressed me out. But now at least I have kind of that, you know, like verifications like this is stuff that's happening because it knows my role. It also knows what I need to

role. It also knows what I need to respond to. So I can like rank it based

respond to. So I can like rank it based on like what's important right now. I

don't know. It definitely it just took away a lot of like the mental burden of like I'm missing out like I'm falling behind in my work or something like that. I don't feel that anymore and

that. I don't feel that anymore and that's I don't know it's a quality of life improvement I guess but definitely just like Chinese schoolwork as ser as finders like you will be surprised that

it is uh gentleman over here uh have you guys like considered going

down design or like UI um tools I mean yeah so design UI tools uh no I mean we did have a collaboration Figma the other day with so there like there

are things happen in there but nothing like built in that I know so far but of course like MCP servers are already pretty useful at time for like big excell

UI right now which is really cool if you ever want to draft like a like a flowchart or something you can just see it in your chat and then you can open it and escal it's it's super nice yeah it's

niceation that gentleman back here yeah uh just building on the Last question, you talked about how you are using co personally. How are you using

it? Is there any example you can give

it? Is there any example you can give that you're using across teams like or collaborating with people?

Uh yes, how are we using co to collaborate with people? Well, we I I don't know if we are collaborate with people just because we don't have like a way to share sessions. Um I mean we definitely are like sharing maybe some

skills or like plugins like but most of our internal plugins are also the external ones that you see on like co-work um that's the main thing I think co-work so far at least for me and what

I've seen across teams is that it's still mainly a personal tool. It's your

personal coworker, but making this more like multiplayer as you said or more for for your co-workers as well is something we just

Yes. Oh, yes ma'am. Right here.

Yes. Oh, yes ma'am. Right here.

Yeah, I was going to say that we're definitely power users as a GTM board because of the connectors and because of all the

plugins. One thing that we brought into

plugins. One thing that we brought into is being lot behind security and it um is these any plans on maybe enabling or

creating maybe some type of where they can open up some of the world to us where we can play around experimenting.

Yeah, we should talk after because maybe questions pass on to the team. I think

I'm not sure for like what your IT like department doesn't like there. Usually

it's around like the MCP tools and stuff with any plugins. Um you do have pretty granular control of what's allowed and what isn't. Like within settings you can

what isn't. Like within settings you can say like oh you allow this action or never allow this action. So we do give people some control but I feel like it's a bigger security question that we should probably talk about. Um or at

least also guys talking to it's a good question.

Yes sir.

Yeah. I was gonna ask something similar because I recently got some backpack from the span company when I suggested cloth code work.

Yeah.

Um so so yeah I guess similar answer. My

other question would be I recently found out that cloth code was a hackathon project. It was like a side project of

project. It was like a side project of course. Are there any other side

course. Are there any other side projects that you guys are sleeping on?

Maybe.

No, nothing. Every time we do something, we're excited about it. We ship fast, so if we want to share something, we share it with the world. Um, but yeah, no, it's definitely funny. Kind of a whole I mean, I'm sure you've seen Boris's interview. It's like it's super fun

interview. It's like it's super fun working with him because yeah, it was just like an internal side project that kind of ruins to one of the biggest pieces of software ever built right now.

Somebody helps you grow props and boards.

My god, you guys are really engaged.

This is awesome. Yes, sir. In the back.

So, large organizations usually uh they have to kind of adhere to certain strategies and frameworks or selling

whatever. um in order to adopt these

whatever. um in order to adopt these kinds of tools uh the use cases have to be very tried and tested and frameworks integral that large. So what are kind of

the suggestions recommendations you have around incorporating those frameworks into these process into these kind of solutions because that becomes really

strategic right that burden that the scale of that in some cases you have thousands of large so what's your problem that

you have I mean I don't know I mean roughly so the question is like as you get mature and you have a larger team. How do you um bake sometimes very complex requirement processes into the

workflows? I mean to some degree you

workflows? I mean to some degree you accomplish some of that with prompt engineering by by stating like these are the exact actions you can take. This is

our voice as our flow that you must always read this URL first before whatever. Um I think you can also get

whatever. Um I think you can also get there to some degree with MCP servers like uh by you know kind of making custom functionality or workflows for your team and then only enabling them.

You could have a private internal MCP server that just ensures that the tools that are called like always produce uniform outputs for everybody. Um, and

then there's also the compliance kind of like policy question which I'm not qualified to answer. I'm not a legal guy.

Rules engine.

Yeah, sort of. And you kind of there's layers I would say to like let it in, right? Where it's like the first and

right? Where it's like the first and cheapest attempt is prompt engineering and then you can start building custom tools or MCP servers too. Um, and then eventually I think we're gonna probably see some kind of guardrails because everyone keeps asking about the need for

that everywhere. Yeah.

that everywhere. Yeah.

Is that roughly what you said the same thing? Okay. Most of this still solved with plugins at the moment and connectors. I feel like that's where

and connectors. I feel like that's where you define your own skills and like just like preferences, but there should probably be a better like tool to make that a bit more like

Yeah, you can drop in work OS offkit and use arbback to tell to I mean seriously there's also you know there's also software plays like you can have an intermediary layer and ensure that agents can only take certain actions

too. So that even if a user asks for

too. So that even if a user asks for something that's not allowed and the agent even listens and goes and does it anyway and the guardrails fail, they can't make that API call.

We can actually just ingest document.

Yeah.

Y suggest and then skills too because to some degree what you're describing is basically a skill. So once you do it one time then you ask now please make this a skill and I'm going to email it and put it in a GitHub repo and we're all going

to use it. Yeah.

Yeah. uh ether back as folks in GM I'm wondering how does the topic use the cos as well as code and are there any creative ways that

you've seen um across that you can share it so I'm not on the GTM team so I'm not entirely sure how many I mean so sorry what was the question oh sorry um how our own GTM team uses

work and the stuff that I've seen them do like phrasing all um no I'm not so I can't really speak on their behalf But just what I've seen from like conversations is that we use

it for pretty much everything. So any

well email of course but also like printing presentations um legal we have good legal plugins. So we at least I mean we don't base all the legal on that.

You also have lawyers. Got it.

Um, yeah. So, no, it's just again it kind of like slowly goes like everywhere. And also whenever we notice

everywhere. And also whenever we notice that people aren't using cloud as much, we're kind of we're just asking like why not? Because this is a gap like

not? Because this is a gap like ultimately we want everyone in a topic to use most of if we don't use it, you know. Um, yeah. So, we're we're using it

know. Um, yeah. So, we're we're using it the same way as most people do. We don't

have any secrets like that. Um, but most of like the new features are based on our own experience. So,

oh, we got to get we're getting the I'm getting the shuffle off the stage. We

got to wrap it up. Let's take maybe one more question. You sir, sorry.

more question. You sir, sorry.

Yep.

Yeah. So, can you also do for conference bring to be able to sort based on topics?

Are the models capable of doing that right now? Yeah, your question is if you

right now? Yeah, your question is if you show claude a whole bunch of different images across categories, is it capable of a seeing them seeing them and b categorizing them correctly? And in my

experience, yes, very very well.

One more.

It's not a question, but I just want to thank you guys for standing up against the Pentagon.

Me, too. Yeah.

No, but but take take the applause back to Yeah. Awesome. Thank you all so much.

to Yeah. Awesome. Thank you all so much.

This is a lot of fun. You're super

engaged. Uh, pizza. Everyone should eat and take pizza and help yourself to drinks in the fridge. Thank you all so much for coming.