CLAUDE COWORK FULL COURSE 3 HOURS For Beginners (2026)

Claude Coowwork is the single most powerful productivity tool available right now for business professionals. And I don't say that lightly because we've trained over 20,000 professionals on AI and automation. And we've tested pretty much every tool out there. Claude Co is different. It's not just another chatbot you type questions into. It's an AI that actually works on your computer. It can open your files, navigate applications, research across sources, process documents, extract data, and it does all of this autonomously while you focus on other things. By the end of this course, you'll know how to set up clawed co-work, delegate real business tasks to it, automate your workflows, and do things that would normally require hiring a new employee or assistant. And you don't need any technical background to follow along. We're going to start from absolute zero, and I'll walk through every single step. Here's what we're going to be covering. We'll start with what Claude Co-work actually is and why it's different from chatbt or regular Claude. Then we'll get it set up on your machine, which takes about 2 minutes. There's no terminal or coding involved. I'll walk you through the interface so you know exactly what you're looking at. From there, we're going to delegate our first real task to Coowwork live so you can see exactly how this works in practice. Then I'll show you how to connect co-work to your tools using connectors. I'll also show you what to do when there's no connector or pre-built integration available. After that, we get into skills, which is probably the single biggest unlock of this whole course. Skills let you train claude to your exact way of working. So, it can handle pretty much any task that you'd normally need to do over and over again without you needing to prompt it each time. From there, I'll teach you about context engineering and why understanding this is what separates people who get average results from people that get incredible results. And then we'll get into some real business work. We'll do four live builds together, including building an agent that runs on a schedule and takes over a task that probably eats hours out of your week. I'll show you how to use co-work on apps that have no API or integration, so nothing is off limits. And I'll walk you through our team's exact setup for how we can use co-work together where all the files that co-work creates is shared across the whole team and not just a bunch of local files on everyone's machine. After that, we'll get into the power features, plugins, live dashboards, projects, sub agents, and more. Then, of course, we'll cover how usage and limits actually work in co-work, as well as my best tips for you to avoid hitting your limit so you can get the most out of your subscription. And we'll finish with the mistakes I see people make over and over again, so you don't have to learn them the hard way. Now, we've got a lot to cover, so let's get into it. Starting with what coowwork actually is. So, what is Claude co-work? And the first important thing that I want to point out here is co-work is more than a regular chatbot. So I guess most of you out there have probably used AI chat bots before. Chat tobt, regular claw chat, or Gemini. Co-work is definitely not that. Co-work is what we would call an AI agent and there is a stark difference between them. With an AI chatbot, when you're chatting with chatbt, Claude or Gemini, what it will do is it will provide an answer. Whereas with an AI agent, it is actually able to take action. This drastically changes the way that you actually need to work when you're using co-work. And that's because when we move from an AI chatbot to an AI agent, we need to move from simply chatting with AI to delegating work to AI. When you chat with AI, you ask a question, get an answer, but then you're the one that actually goes needs to go and do that work. When you delegate work to AI, you just describe the outcome. Describe what needs to be happen. describe the job that you're trying to do. AI then goes and in this case Claude called coowwork goes and plans the work. It'll use its context, the files and folders and tools that you give it access to. It's going to ask you for approval when needed and that co-work actually produces the actual output. And this is a very important point that I want you to keep in your mind throughout this whole course: is the shift in mindset.

Now, another way to think about this is comparing the three different Claude solutions that you have on offer, which is Claude chat, Claude Co-work, and Claude Code. And this is a question I get asked a lot when we're giving our company trainings is what's the difference between each one and when should I use each one? So, starting with Claude chat and that is just a regular AI chatbot and what we would call a thinking partner. You ask it something, it gives you an answer, but then you need to actually go and do something with that answer. Now, this is still great for brainstorming, research, thinking through a decision, but the actual work that needs to get done, that's still on you. And here's where co-work is different. It's what we call a work partner, meaning Claw doesn't just give you the answer. It takes the action. It can process tasks, work on your documents, run multi-step workflows across your actual tools. And then, of course, finally, we have Claude Code. And this is what we'd refer to as a build partner. Claude code is a tool aimed at developers and it's best suited at building apps and tools, writing custom automation scripts or debugging and fixing code inside a codebase. Now, one important thing to point out is both Claude Co-work and Claude Code are agentic platforms. They're both AI agents as we just described, but they're just designed with different user groups in mind. And we can see that clearly looking at Anthropic's own description of what Claude Co-work is. Coerc is Anthropic's agentic AI system for knowledge work. It runs on desktop, connects to local files and applications, completes multi-step tasks from start to finish. The user defines the goal, co-work figures out how to get there, and how it's able to do this are six main capabilities that I just want to point out quickly before we get into actually setting this up on your machine. So, here's what I think makes co-work so useful. I'll just touch on these quickly. As we've already seen there in that definition, it can work locally on your computer and access local files. It also has the ability to write and run code. So even though co-work is focused on knowledge work, the ability to write and run code is incredibly useful for knowledge work. The one difference between claude co-work and code here is that all of this code that runs happens inside a sandbox on a virtual machine on your computer and not directly on your computer like claude code, which just makes it a little bit more secure. Additionally to this, we can connect co-work directly to the tools that you use each and every day. Think Gmail, Slack, notion, Salesforce, HubSpot, whatever tools you're actually using in your business. We can also create skills and plugins to create reusable workflows. And we'll touch on all of these throughout the chapters coming up. But two that I just want to quickly highlight very quickly are what is known as an agent loop. This is a very important concept that I think you should have in mind as we start setting up co-work and giving it real tasks. So what is an agent loop? Whenever you give co-work a task, as I said, co-work is an AI agent. And what that agent does is figures out what needs to be done in order for that task to get completed. After it figures that out and makes it plan, it actually goes ahead and takes action. That could be editing a file, pulling in some data, writing a document. And here's where the loop part comes in. After it actually takes the action, very importantly, and this is something that chatbots do not do, the agent actually checks its own work. After it checks its work, it decides what to do next. That could be repeating the process, maybe having a second iteration or moving on to the next step in the plan. And that is this agentic loop that gives co-work the ability to do real tasks. The next very important factor that makes it super super useful are sub aents. And I just want to quickly touch on why sub aents or parallel agents are so useful in co-work. If we take a look at this example prompt where I'm asking co-work to research five competitors and for each competitor I want it to do very specific things in its task without the ability of having sub agents. What co-work would need to do is go and do all of this research for competitor number one, then go and do all of this research for competitor number two, then competitor number three, and so on and so forth. And what typically happens in a situation like this is you'll probably get a very good result for competitor number one. But as it starts working through, the AI can tend to get a little bit lazy and you'll also notice the results that you're getting for three, four, and five is drastically worse than the first couple. With parallel agents, what co-work can actually do is spin up several versions of itself. So one agent will just do all of this research for competitor number one, while another agent does it for number two. So he would have five separate agents working at the same time and allows it to a get far better results, but also you're not biased in your work. When the agent or sub agent is researching competitor number two, it is not influenced at all by the research happening in competitor number one. And then again, not to get into too much specifics right now, all of the sub agents basically report back to the main agent on the work that was done and hand that over. We get a lot of gains and efficiencies from this that I'll be explaining throughout the course.

All right, let's actually get co-work set up on your computer. And as I said, the process is pretty simple. It should take less than a couple of minutes, and you don't need to worry about terminals, coding, or anything like that. The only thing you do need to be aware of, however, is that in order to use Claude Co-work, this is only available on the paid Claude plan. So, if you're just on a free Claude plan, this does not include Claude Co-work or Claude Code. This starts on the pro plan if you're an individual. So, if you want to use monthly billing, that's 20 bucks a month. Or if you set it up for an annual subscription, that's just $17 a month. Also, if your company is using Claude, whether that's on the team or the enterprise plan, both of these have access to co-work. So, the very first thing you need to do if you do not already have a paid claude subscription, make sure you go ahead and sign up before continuing. Now, I know that this obviously does involve a monthly cost and people are in different situations, but I have to say this is 100% worth it when you want to start using this for boosting your productivity or your productivity at work, especially when you compare this price to other subscriptions you might have like Netflix or Spotify. I think this pays for itself 10 times over. Now, in a later chapter, I will be covering in detail around usage and limits. For now, if you want to get started, I recommend just get started on the Pro and then we'll speak about this difference between Pro and Max a little later on. Now, once you've got your paid account set up, the next thing you need to do is actually download the Claude desktop app because this is where Cowwork lives. This is not something that's available on the web version of Claude. You do need to have the desktop app installed. So, what I want you to do is head over to claude.com/d download. I'll leave a link in the description below. And then simply download the desktop app for whatever operating system you're using. So, in my case, I'm using a Mac, so I'm just going to download for Mac OS. And then just go ahead and install it like you would any other app. Now, once that's installed and you've opened up the Claude desktop app, you should land here in the chat interface. This is the exact same as using Claude chat on the web. But to confirm that you now have access to co-work, what you should see is at least one other tab. So, one or two other tabs here. You can see here co-work. And when we click on the co-work tab, you should land here on a new task. And that means that you do have access. Now, if you do not see co-work here and you're on a team or enterprise plan, that means that it needs to be enabled for your organization by your claude admin. If you happen to be a claude admin, I can also show you those settings. So, you need to go to organization settings. Then underneath products, select co-work. And here you'll have the toggle enable for your organization. So, now that you have co-work set up on your computer, before we go and give it its first real task, I just want to walk you through the interface so you know where everything lives. And the first thing to point out is a little bit of a change in terminology when we move from using the regular clawed chat where every time you start a new conversation these are known as chats in co-work because you're delegating work to an AI everything is referred to as a task. So if you want to give a new task to co-work you can click here on new task and here you have your new task interface. Obviously the conversation box where you can provide the uh written instructions the prompt to co-work to work on the task. Here we have our model selector. I really recommend when you're getting started and trying to figure out co-work see what it's capable of, maybe what it's not capable of. Always use the best model. As I said towards the end of the course, I'll be sharing my tips and tricks to make sure you're not burning through your usage like crazy. In my case, I pretty much exclusively work with Opus in Co-work here and don't really have any issues when it comes to limits. Now a main area of confusion in this interface here you see we have this plus button and you can add files or folders and then we also have the option here work in a project when it comes to claude coowwork every task you give it belongs in a dedicated folder and we're going to see this in a second in our first real task but in general always make sure that you have a folder or a project selected and in the case that you haven't got one set up yet if you scroll down to the bottom you can simply from here create a new project, which we're going to do in a second, or you can choose an existing folder on your computer. So, you just open that up and find the exact folder that you want to work with. On the left hand side here, we also have this left hand menu where you can see here projects, schedule tasks, live artifacts, dispatch, and customize. In customize is where you're going to be able to set up all of your skills and connectors that we're going to touch on in the upcoming chapters. And also you can basically collapse and open up this sidebar when you want to have a more focused view on the task that you're working on. Then all of the previous tasks you've already worked on, you can find here in the recents section. You can either scroll through here or click view all to find this view of all of your tasks. You can archive tasks if you want them uh hidden. And you can use the search bar to find an old task that maybe was done way back. Let me also show you quickly um the interface when you're working on an existing task. So here was a recent task I worked on. What again what is going to be a little bit different when it comes to chat is this right hand side here. So first of all we have the progress bar and in the progress bar is going to be where Claude lists its exact plan of what it wants to work on. So when you give it a task it's going to create a list here on the right hand side especially for bigger tasks. And then you'll basically be able to see it tick these items off one by one. Any files that Claude has access to inside that working folder or project folder that you provided, you'll see here. And you can always access that folder in one click. And at the bottom here, you'll see what tools Claude has connected to while working on this task and any skills that it's used. All right. Now, let's get straight into it and give Co-work its first real task. is for this first demo. What I want to highlight or showcase is the power of letting co-work inside a folder on your actual machine. So here the use case is it's uh tax time at the end of the year and we've collected all of our invoices from the previous year. So here inside this 2025 invoices folder, I have more than 60 different invoices that I now need to share with my accountant. And my accountant needs this in a very certain format. It needs to have a full CSV with all of the details from all of our invoices. Now, if you actually want to try this one uh alongside with me over in our community, underneath this video, you'll be able to download all of these uh mock invoices that I'll be using um for this uh first task as well as all of the prompts used. So, check the link in the description um and join our community and then you can follow along with all of the resources that I'm sharing in this course. So to hand over this task to co-work, what I want to do is make sure I'm in the new task tab. And the very first thing I need to do is give Co-work access to that folder where all of the invoices live. Now, I touched on this briefly um before, but here when you work in a project or inside a folder, you're actually giving Co-work the ability to read, write, edit, delete files all within that folder. By default, co-work is not allowed to just go and look in your downloads folder or look in your desktops folder. If I ask it here to, hey, show me all files in my downloads folder unless I actually what is known as mount that folder as the working folder, it cannot access it. So, in this case, I have my invoices in a folder. I'm going to choose to work in that folder. And then I just head over to my desktop and select that folder with all of the invoices. Now we'll get a message coming up saying do we give Claude permission to change the files inside this folder. For this task I'm going to select always allow. Obviously you do need to be careful. So if for instance the files are very let's say hard to get and you will not be able to download them again maybe create a duplicate folder because really if you prompt it uh the wrong way Claude is able to delete files also inside that folder. In my case, I'm going to always allow. And this should be the way that you always provide Claude with the files and folders that it needs to work with. There is another option here when we click on the plus button, add files or folders. These are files that you give with the prompt. And the example that I'll give here is let's say you have Claude Co working on a design and you want to share a screenshot that of either a design mistake it made or a website that you want it to emulate. That is the time that you would actually add here what is known as attachments. And what happens is Claude then immediately views those images and uses it as part of the prompt. If you want, however, Claude to actually work on some files, read them, edit them, update them, that always needs to live in the mounted folder or in the working folder. So now that Claude has access to that folder, the next thing I need to do is give Claude its task. I've prepared this prompt here, and there's a few parts to it. First of all, what I always like doing is when I give Claude access to a brand new folder is tell it what's actually inside. All right. So, here I'm saying in this folder are all of my invoices from 2025. And now I move into the task at hand. So, I need you to process these so I can send them to my accountant. I would like you to read each invoice and extract the following data points. So, here I'm listing exactly what data I want it to extract from all of the invoice PDFs. And now I move into what files it should actually create and what changes it should make. So then create a CSV called invoices 2025 CSV where you store all of this information as well as the file name for each invoice. After you've extracted the information, move each invoice PDF to subfolders based on the year and month. So imagine this is the format that my accountant expects it. This is the task that I'm now providing to co-work. So let me send this one off. Now the very first thing that co-work is doing, you might have seen the first words they're starting up. For every new task you create, co-work creates its very own virtual machine on your computer. And now any work that it does happens inside that virtual machine inside the sandbox and the only other area it can access is the mounted folder that you have provided. So now what has co-work done? It started by laying out its plan. So, what it's going to do, it's going to list every single invoice inside the folder, extract the invoice data, create the invoices 2025 CSV, and then move everything into the subfolders, verify the CSV and folder structure. So, this is that agent loop that I was mentioning at the start after it's finished its work. It's not just going to say, "Hey, I'm done." It's going to verify and check its own work. Now what you could do at any moment is if it starts making its plan and you're not happy with it, you can basically hit this stop button and then send your next instruction. Now as co-work is actually working through the task, you can always click on this working text and you can see every single thing it is doing. So in this case, it even created its very own Python script to actually pass the data from these invoices. And literally a couple of minutes later, we can see Claude is already finished with this task and it's giving us a summary of what it did. It extracted all the data from the different invoice PDFs. It even points out here some interesting things that it found. So, some file names were misleading or had incorrect names that didn't match the invoice date. So, you can see just the level of detail it goes to. It gives us a summary here of how many invoices we get per month. And then, obviously, we have our invoices CSV table that we can preview here directly in Co-work. I do have to say that the preview is not the nicest here on the inside co-work itself because you can see some of the data sort of crossing over columns. What I would recommend if you do want to look at the final data is opening this up in Excel or Google sheet. More importantly, we can now head over back to that invoices folder. So I can either go back into my finder or at any moment whenever you are working inside a folder, we can simply jump to it by clicking go to folder. And now we see something very different. All of those unorganized invoices are now perfectly, let's say, categorized by um the month of the year that they belong to, which was our exact instruction. And this happened in uh a matter of minutes. And here maybe we can see a little bit cleaner that um summary CSV that we asked it to create. I really think this is just a a great example of the power of co-work. If you wanted to do this task manually yourself, it would take a ridiculous amount of time and it's an absolutely mundane task that no one really wants to do. And here we've been able to delegate that off to co-work. So now you've just seen co-work on files in a folder on your local machine, but most of your work probably doesn't live in files alone. It lives in the different tools that we use each and every day. And typically this is on the cloud. So let me show you now how to give co-work access to where your work actually happens by showing you how to connect co-work to your different tools. And in co-work there's actually three different ways that we can interact with the different tools that we use. First of all, using their connectors feature. This is by far the uh main way to do this. So co-work can access your tools via builtin integrations or custom MCP. Now if there are no connectors available for the tool that you're working with, we have a few other options. So there's also code co-work can run scripts on its virtual machine to connect via API or CLI. This gets a little bit more technical. And then finally, also browser automation. And I'd put computer use, which is a newer feature into this category as well. Claude can actually open up your browser. We'll touch on this a little bit later. It can open up your browser, navigate pages, click on buttons, and and extract data. And this is very, very useful when existing connector or API doesn't exist for the tool that you're trying to connect with co-work. But as I said, connectors, this first category here is by far the simplest, the one I would recommend to get started uh with and that's what we're going to be focusing on in this chapter. All right, so let me show you how to connect your tools with Claude Co-work. So what you want to do is head over on the left hand side to customize and then we see here on the top connectors. So connectors are the different apps that we can connect to Claude and to Claude uh co-work. You can see I have a bunch already set up here. If you're doing this for the first time, this will most likely be empty. So, what you want to do is click on this plus button to add a connector and then browse connectors. And here you'll land on the full list of the hundreds or maybe thousands of different uh in-built connectors that are set up. So, I would highly recommend if you want to connect a tool to Claude, definitely start with connectors first before you go down the API or browser automation approach that I'm going to show a little bit later on because this just makes it super super easy. Now, one thing I need to call out, if you are on a team plan, the connectors that you have access to is going to depend on what your admin allows. So, if you search for a connector, in my case, I am a uh admin. So, let's say I search for HubSpot. If HubSpot, let's say, was not added to your account, you'll see a little message here that you need to contact your admin and ask them to enable it. So, in my case, I'm an admin, so I see every single uh tool here. Let me just set one up for you um so you can see the full process. So, I'm going to connect my uh Gmail. When you um select the tool that you want to work with, you can also click on this little cog symbol. And you also get a little bit more information. You can go and check out the documentation and also what tools this uh app, this connector allows. So, if you are in an individual plan, most likely at this level here, you'll already be able to directly uh connect to this uh Gmail connector. In my case, because I'm on a team plan, it's basically a two-step process. So here if I've already ensured that Gmail is available to my whole organization and now if me individually if I want to connect the Gmail connector I'm going to scroll down here to the not connected connectors and these are all of the connectors that we've enabled for our uh team plan. I can just click on Gmail and then connect. Now what's going to happen is it's going to ask you to log in and give Claude access to your specific account. So here I'm going to select my email account. And this is the way that most of the connectors are set up. They're using something known as oorthth. So that when you connect this app to your claude, you'll only have access to the things that you should have access to uh in general when you're logged in. So let's say you want to connect to notion, you're not going to be able to or Claude's not going to be able to see notion pages that your user that you're logging in with doesn't also have access to it. It's the same with Gmail. Obviously, by granting Claude access, it doesn't automatically see Gmail uh email messages from my colleagues. It's only going to be able to interact with the email address that I'm connecting here. Once I've gone through that process, I just need to click open in Claude. It's giving me the verification that it's connected. This is going to reload in the back here. And now we see Gmail has moved into our connected connectors. And now the most important thing to understand about connectors is the ability that you can do with each and every tool is dependent on the tool you're using. So in my case with Gmail, we see here we get a bunch of and these are known as tools. So tools are basically the abilities that Claude um will have to do something in that specific tool. So in this case for email I have some readonly tools. Readon means it's going to extract some information from the tool in this case from Gmail. So, it's able to retrieve a specific email thread. It's able to search email threads. And it's able to list user labels. We then also have write or delete actions or write or delete tools. So, in this case, it can create a new email draft. It can create a new label, delete a label, list draft emails, and modify existing labels. And for each individual action or each individual tool that this connector is able to do, we can basically set some permissions. So we can set it for the whole category. Like if I collapse this, you'll see that we have three readonly tools and five write to tools. And by default for every single readonly tool, I want to say that Claude can is always allowed to do this and it doesn't need to ask for my permission. We have three options here. We can either always allow. So this is where Claude can just use this action without asking for permission up front. We can also set it to needs approval, which means before it wants to go and search our emails. It's going to ask permission, say, "Hey, based on what I've detected, I want to search your emails. Am I allowed to do this?" If we wanted to set that up, I just need to toggle this to needs approval. And now you can see here that the category has changed to custom, but I can always move this back to always allow, and it's going to update everything. And then the third and final one is if if there's ever an action that you never want Claude to do under any uh circumstances, we can set this to blocked. So typically the way that I like to set it up um at least to begin with is readonly tools by default I'm going to leave as always allow. And any write or delete tools I'm going to set to needs approval. as I start using it uh more and more and maybe when I we create some specific skills around this, there's going to be some actions that I would typically then toggle over to always allow. And this is something that you really need to decide based on like let's say how dangerous um it's going to be if Claude gets this wrong. In my case, creating a new label in my inbox. Maybe it's annoying if Claude just goes and creates some 20 labels, but it's not really going to have any big uh impacts. Where you always want to leave this as needs approval is for instance if it's going to go ahead and send out an email directly. Now another quick tip here is that these connectors are getting updated constantly. This is a very new technology. All of these use something known as MCP under the hood. MCP is very new. A lot of tools are trying to quickly get their connectors out so that people can use them in tools like Claude and Chat GBT. And they're also updating what they can do constantly. um and it doesn't actually automatically update. So if you hear for instance that Gmail has some new uh abilities and you log into Claude and don't see it updated, if you click on the three dots next to a connector, you have this option refresh tools list. In my case, if I uh click this one and we see the tools list refreshed and we actually have even here now for Gmail some new actions. So we now have nine instead of I think it was five previously write or delete tools. So now we also have the ability to uh add labels specifically to a message. So I guess people are using it to automate um cleaning and sorting out their inbox which is very very cool. Now something that I'm seeing uh more and more with connectors is there sort of two different approaches when it comes to different apps creating their connectors. There's this approach here that Gmail have done where they create tools that are very very specific. So you can clearly see that every single tool has a specific action here. List user labels, search email threads, retrieve a specific email thread, remove labels, add labels. Every single tool is very very specific. Now there's another category that I'm seeing more and more. And one of the ones that we have um here is customer.io, which is our email marketing tool we use. So you can see here, let Claude work directly with customer workspace to create segments, inspect user profiles, search for customers, analyze attributes, and so on and so forth. What's different about the way that this connector is set up is it doesn't really have, if we look here at the right tools, it doesn't really have very specific uh tools. What I'd expect here, like create a marketing email, send a marketing email, create a segment, create a marketing campaign, right? This is all the things that an email marketing tool should have. instead that the way that this connector is set up is they pretty much allow Claude to do anything that is possible via their API. So here it has two options, two actions to write data. It can write to the API or it can delete from the API. And then what's that coupled with is some read only tools. And typically what they've done is a little bit meta, but they've created a tool so that Claude knows what's even possible with this tool. So for instance, it can browse the API schema. Meaning that when I ask Claude what's which we're going to do in a second, what's possible with customer, it's going to probably call these three skills here that it can browse the API schema, probably also read from the API. And then what the tool has done is they've created their own skills for Claude to be able to understand exactly um how to then make those API calls. So this is a little bit hard to understand. And what I'd recommend in this case is let's just go back over to a new task. And also at any time if you want to make sure like let's say if you want to disable a connector in a specific task, what you can always do here is click on this plus button, go down to the connectors and you can also disable it. So let's say if I had a second uh email marketing tool and I didn't want Claude to get confused, I could just go ahead and disable that one. So let's say if I toggle on and off air table that's going to have it disabled for the upcoming task. In my case I'm going to say I've just added the following connector. Can you please take a look at the available tools and let me know all of the possibilities that you now have to interact with customer IO and then also list out some based on our work together some potential use cases where I could leverage the customer IO connector. And this is something that I really recommend uh you do whenever you connect a new tool because chances are there's things that are possible that you have not thought about yet. So let's send this one off. And we can see the very first thing that co-work is doing. It looks a little bit hard to decipher here, but this is literally listing all of the MCP tools available to it. I think it's only doing it with maximum 10 at a time. And now here we get our answer. The customer connector gives me eight tools and it lists them out here. here. And you can see there's an even more in-depth explanation than we get uh in the settings. So, for instance, for the skills list and skills read, this is customizo's owned packaged playbooks for common workflows. And what's interesting here is I know that it hasn't actually called this tool yet. So, it actually hasn't done that. But even without doing it, just by looking at the different options and based what it knows, it's already given me a bunch of potential use cases, even syncing from one tool to another, which is quite interesting. So, what I'm going to do because I know um that it hasn't actually called this tool yet, I'm going to ask it to do that. Can you please call the I'll just paste this one in the CIO skills list tool and tell me everything it returns. I want to know all the playbooks. And this is also something you can do even with tools like Gmail where they have more specific names for their tools. actually just having a brainstorming session with co-work about just exactly what's possible with that connector and it will give you a very good idea about how to get started. And now here's a great example. When it comes to actually one of these connectors using one of its tools, so doing one of these actions, it's going to ask me for permission based on those settings that I have set up in the connector settings. So in this case, I guess everything was set to needs approval. So Claude is asking me is it um even though I specifically asked it to do it here, it still needs approval to use this list available uh skills tool. I have a few different options. I can allow once just in this chat or I can say always allow and that's going to mean whenever we're working in this task, it's going to be always allowed to use this tool. So then we go here's everything uh returned. There's four top level skill bundles um each with a set of subfiles you can drill into. So, I won't spend too much time like going through each of these in detail because maybe if you're not familiar with email marketing, some of these won't really speak to you, but I just hope that you get an understanding of what I'm trying to show here in terms of connecting the tool to Claude and then asking how it can help you in your workflow. Also, especially if you have a specific idea in mind about what you're trying to achieve, ask Claude if this is something that is possible. Now, we're going to be using connectors extensively through the different live builds that we do, and I'll cover all the best practices. But for now, let me just give you a quick demo with that new Gmail connector that we set up. So, I'm going to create a new task. In this case, it's just a demo. I'm not going to be setting a working folder, but remember, you should always set uh a working folder. And I'm just going to ask co-work, can you please take a look at all the emails I have received today? And let me know if there's anything urgent and important that I need to act on right now. Urgent and important is only when the email is addressed to me and I'm the one that needs to act. Ignore any sort of marketing or notifications that we get from different SAS tools. Now, something interesting with connectors is we can't tag them specifically. What you're going to see in a second is for instance when we set up skills we can do a forward slash and basically directly trigger any of the skills uh that we have set up we can also mention I believe um specific projects that we have uh set up or specific folders with the at symbol um but it's not the same for connectors so for instance if I do forward/gmail we don't see that Gmail connector what's basically going to happen is Claude is going to understand based on the context of this message that it needs to use its clawed connector or you could be here even more um specific. So I could say all the emails in Gmail but in this case it wouldn't have made a difference. Let's send this one off. And we can see it's already detected it needs to use the Gmail connector and what it's doing it's searching for email threads. Remember in my Gmail settings I said whenever it wants to read data it can always do this without asking for my permission. So that's what we're seeing here. It started actually searching through uh my email. What's really nice is when it does this tool call, which is what the the name of this is, you can actually click into it and see exactly how it is structuring that request. And this is very very important because if we take a look here, what it's querying is all the emails sent to me specifically newer than one day that are not in draft and are not in category promotions, category social, category updates, or category forums. And very importantly, it's limiting the page size to 50. And this is something that happens more and more and sometimes you might need to instruct. I think what happens in most cases is let's say if I received more than 50 emails today, Claude's probably going to detect that from its um first response and then run another call for the next um load of emails. But in many cases, it doesn't do this. So you really need to be keeping an eye on exactly how Claude is structuring the calls that it makes to the different tools to make sure that it's getting the accurate data from the tools cuz imagine maybe in my case I wanted to also include promotional emails in this uh search query. I could basically just tell it to go and update this. And we also have a similar thing with the response that we can see here all the emails that we get back. So, let's take a look at Claude's response saying, "I went through everything that hit your inbox today. Good news. Nothing urgent that needs your action right now." And it gives me a rundown of the different categories. So, there's some automated notifications that I might need to be aware of. Some uh cold sales pitches, safe to ignore unless you want to reply and some threads that you're CCD in. But it mentions here that this is the closest thing this um hackathon preparation where we're working with one of our customer but basically clarifies that my colleague Audrey is the one handling this. Now one last thing that I want to cover on connectors right now before we move on to the next chapter and that's the fact that of course while Claude has a bunch of these builtin uh connectors already added here chances are that the exact tool that you want to connect to co-work might not be available yet. That's also been the case for us. For example, the uh tool we use to run our community does not have a dedicated uh builtin connector listed here um right now. But before you give up, the one thing I want you to do is head over to Google and just type in the name of the tool that you want to connect. So in our case, we run our community on Circle. So I'm just going to type in circle.so and then look for MCP server. And what we can see from the results that there's a couple of articles coming up. I would always recommend looking at the developer documentation. So in this case, Circle does already have an MCP server. MCP stands

For model context protocol, and it's actually the technology that all of the clawed connectors are built on. So if the tool that you want to connect does have an MCP server, we do have another option here in the connectors. When we click plus, instead of um looking, clicking on browse, we can select add custom connector, and we'll need to provide two things. So first of all, the name. So in this case, it's going to be easy. I'm just going to call it circle, which is the name of the tool. And then it's asking for a remote MCP server URL. And that we're going to grab exactly from the documentation of the tool. And you can see here, in the case of Circle, it is the first thing listed here in the setup, the connection URL. So all I'm going to do is copy this one, head back over to co-work, paste in that URL, and hit add. In my case, I already have it set up. So it's giving me this warning: server with this URL already exists, but it's that simple. Once you have that set up, um, you can basically go ahead and connect like you would. So, I'll just disconnect this one so you can see uh the process. So, now um your custom connector will be appearing here. And all you need to do is hit connect and follow the exact steps like we did before. So, in my case, we're going to give it full access to our Circle community. And then once you're connected, it is listed just like any other connector. There really is not much uh difference between the built-in connectors that Claude offer. I guess they've just maybe gone through some extra verification or validation from the Anthropic team. I expect more and more tools to be listed here. But really, it's just so simple to connect any MCP server. And as you'll see, then, as soon as it's connected, the functionality is identical. We have all of the tools that that MCP server allows. And we have access to the same permission settings like we do a built-in tool.

All right, now that we've connected co-work to our tool so that it can actually do real work for us, next we need to actually teach it how we want that work to be done. And that is where skills come in, which is what we're going to be covering in this chapter. And I really can't understate this enough. Skills are probably the biggest unlock for getting the most out of co-work. So, let's get right into it and talk about what skills actually are. So, what is a skill? And this can also be referred to as an agent skill. A skill is a reusable set of instructions that tells an agent, in this case Claude Co-work, exactly how to perform a specific task. And now, before I show you exactly how skills work and how to set them up and how to build them for yourself, let's take a look at a couple of skills in action. So, the first skill I want to show you is one that I've set up to help me create the assets for our 9x live workshops. Basically, every single week we host a free 90-minute workshop on the latest in AI and automation. So, you can see we have a bunch recently covering Claude co-work, covering open claw, covering Claude Co-work for marketing. And of course, for every single event, we need to create some assets to help us actually promote the workshop. Now, previously, we were doing this manually in Figma, but thanks to Claude Co-work, we've been able to hand this over to a skill. So, let me show you how it works. In co-work, I'm going to create a new task. And then if I type forward slash, we see all the list of skills that I have set up. And we see the top one, 9x live create assets. Now, all I need to do is tell it the name of the event. So, let's say the next event that we're going to run is context engineering in Claude Co-work. This is actually what you're going to learn in the next chapter of this video. And I can just say that I am the host and send this one off. And we can see co-work's gotten to work. I'll generate all four assets for context engineering in Claude Co-work. And just like that, it's already opened them up. We have the four different images here in the different formats um we need. So we need one for YouTube, one for email, one for Luma. So for instance, Luma is a 1x1. We have YouTube a 16x9, and then the banner for our community in circle. And because co-work's working directly on my computer, I also get all of those images saved in the working folder. Now, this skill doesn't actually use any connectors uh whatsoever. It's purely using some Python scripts to come up with these image uh layouts. But let me now show you another skill where connectors are involved. So, here's a task that I actually need to do. We've drafted a couple of emails in Notion for our upcoming uh Claude Co-work cohort that we're running, which is a three-week live uh program. And I've pasted in the notion pages where we actually have the email copy drafted. And then I've told it to use this create marketing email skill. I'll send this one off. And what this one's going to do, it's actually going to tell co-work to go ahead and create that email directly in our email tool, which is customer.io, using our exact email template. And we can see Co-work is all done with the task. It's completed its progress here. What you'll notice is it's used two different connectors. It obviously realized it should use uh notion because I shared with it notion pages. But then inside this create marketing email skill, we're instructing it to use the customer.io connector. And then here we have those two emails created. And in one click, I can just open them up. And here we see that email actually created inside customer.io using our 9x email template. In the notion, there was a placeholder to add an image. So it's added that one there for us just to connect. And the rest of it is ready to go. All the links and everything working. Now, this skill is only just one of the many that we have set up for ourselves at 9X to enable us to work a whole lot more efficiently. Previously, only a couple of us had access to customer io that knew how to use it, how to set up branded marketing templates. So, whenever anyone in the team needed to create an email, we would be the bottleneck. Now, we've basically been able to create this skill in co-work and anyone in the team can create on-brand emails.

So now that you've seen a couple of examples of some skills in action, skills that I personally created and are using very, very often, I just want to touch on the main problem that skills solve. So maybe some of you are used to working in chat GBT and maybe built some custom GBTs, or if you're working in the regular clawed chat, you were using projects, and the one main problem that skills solve is previously in those, using those systems, in order to give that custom GBT or that claw project multiple capabilities. So in this example here, if we had a content marketing agent as maybe a custom GBT or a project, we'd have a very, very long system prompt telling it, "Hey, here are all the guidelines for writing good LinkedIn posts. Here's all the guidelines for writing blog articles. Here's all the guidelines for creating newsletters. Here's all the guidelines for writing website copy." Basically, anything that content marketing agent would need to know would need to all be put in the system prompt. And this has one massive issue: is the fact that all of this takes up a lot of space in the context window. If this is the first time you're hearing about the context window, don't worry. I'm going to be covering that in the next chapter. But really, what this means is that regardless of what task it was working on, so let's say I was asking it to write some blog articles for me, it would still have all the knowledge around the guidelines for writing LinkedIn posts at the same time. It would have the guidelines about writing newsletters. And this can eventually cause confusion, and it can be mixing up instructions between the different areas. And in my case, this is something I did actually have set up for myself as a claude uh project. You can see here this very, very long system prompt. We obviously have our brand uh guidelines, and then we have one section on how to write LinkedIn posts. If we scroll down, we have another section on how to write blog posts, and so on. So you can see all of this always needs to be in the memory of Claude when it's working on its task. So, how are skills different, and how do they solve this main problem? And an analogy I always like to use is comparing it to like a filing cabinet. So, instead of having all of those long instructions in its memory at any one time, it basically just has access to this filing cabinet where it knows, okay, I know how to write LinkedIn posts, I know how to write blog articles, I know how to write newsletters, website copy, Instagram captions, PR releases, whatever a content marketing agent would need to know. And only at the time that it needs to write a LinkedIn post. So, let's say I asked this agent, hey, write me a LinkedIn post on this topic. It would just go and open up that LinkedIn post folder, read the instructions, get the best best practices on how LinkedIn posts should be written in our tone of voice, and then execute the task. And the reason that it's able to do this is because under the hood, skills are documented on three different layers. First of all, we have metadata. So this is just the name of the skill and a brief description about what that skill does. Then we have the instructions, and this is a long text file that is the exact step-by-step workflow that the skill should follow. And then finally we have resources. So these can be other files and documents that the skills needs to use when it executes a workflow. So let me show you what these three layers look like in one of those skills that I just demoed. If we head over to co-work, we can find our skills in the customize section, and then we click on skills, and we can take a look at that 9x live create asset skill, which is the one creating the different assets for our weekly workshops. So first of all, we have layer number one, the metadata. This is the name of the skill. So in our case, it's called 9x live create assets. And then here you can see a short description about what the skill does. And the reason why skills are so much more efficient when compared with cramming everything in the system prompt is that these three layers are loaded, are read by Claude at different times. So we have the three different layers, and then we have when Claude sees them. And so the first layer, metadata, this is always in the context. So this is like that tab of the filing cabinet I was explaining. For all of the skills that we have set up, Claude always knows the name and the description of that skill so it can decide exactly when it should use it. Now based on this metadata, if Claude decides to actually use this skill, or we can also directly trigger it using those slash commands that I sent earlier, then comes layer number two, the step-by-step instructions. And these are obviously read when the skill is triggered. If we look at our 9x live create assets, these are all in this markdown file. So every skill has a skill.md, and these contains those step-by-step instructions. Here you can see, and this is just like a very, very long prompt. It is just text in markdown format. We can toggle here between the nicely like formatted view, or if we flip flick over, we can see the exact markdown file. Now don't worry if this already looks a little bit daunting, thinking, okay, how the hell am I going to set all this up and write out all of these skills in such detail. Don't worry, you don't need to do any of this. In a second, I'm going to show you the simple, simple, super easy process for creating your own skills so that you can have Claude trained on any workflow you need. But first, let's look at the third layer. And here we have resources. So these are read by Claude when the skill needs it. And it's in this order that makes skills so efficient. So we have the metadata always read. As soon as Claude realized, hey, I need to use a skill, it's going to read those instructions. And both based on those instructions, it will tell Claude to use any necessary resources. And in our case for the 9x live create assets skill, we have a few different resources. First of all, we have assets. We have some background images that Claude needs to use for creating the designs. We have the font that Claude needs to use. And we have, of course, our 9X logo, as well as the images of myself and my co-founder. So when I tell Claude, "Hey, Yan is going to be the host," it knows it should use my image. If I tell Claude Alex is going to be hosting the workshop, then it will use Alex's images. And obviously, it only accesses the files that it actually needs in the use case. So, when I'm hosting, it doesn't need to do anything with Alex or Piv's image. Next up, we also have some saved scripts. And this is another reason why skills are so efficient. Typically, when Claude is going to do some work, it is going to write some code to execute the task. And all that is baked then into the instructions is how that code should be used. So in our case, we have a different Python script because we're using a Python image library to create the three different types of assets that we need.

So now that we've seen under the hood of how skills are set up and how they actually work, now let's actually get into the fun stuff of creating some skills and adding some skills to our co-work account. One thing I need to address before we get into the building. Obviously, there are skills being shared left, right, and center all over uh the internet. We're also sharing several skills in our community that people can import into their co-work account. I think these are brilliant for understanding what's possible, seeing how others are tackling certain workflows. But the main leverage that you're going to create for you and your team, especially if you're working on co-work at your company, is teaching it your way of working, teaching it your company's IP, how your things should be done. So, even if you're going and grabbing some skills from the internet, I always recommend customizing it and making it your own. Now, if there are skills that you do want to import uh into your account, just go on the customize section, head to skills, and then at the top, you can click on this plus button to add a skill. And you see you have two different options here. We can browse skills. So, there's actually a bunch of skills um that are built in, provided by the Anthropic team. One very, very important skill that I want everyone to set up. This is the one skill that I say you should absolutely take uh and import a skill that someone else has set up, and that is anthropic's skill creator. It should be enabled by default, but if it is not, do make sure you go over to your skills and enable this skill creator, because that is going to be the skill that we use to create every skill that we're going to be setting up. I know it sounds a little bit meta. We're going to be using a skill to create skills, but it is by far the most important. Now, apart from the anthropic provided skills, if you happen to download a skill from the internet, you can just go here and create a skill and upload a skill. But I really cannot emphasize enough: by far the best skills that you will be setting up in your co-work account will be the ones that you create for your own from scratch. So, let me show you exactly how to do it.

So, here's my very, very simple process for creating your own co-work skills. And it starts first of all by doing the task together with Claude step by step. And what I mean by this is a lot of people head into co-work. And as I just showed you before, there is the inbuilt skill creator tool that allows us to build skills. And let's say you want to teach Claude, you want to build a skill for creating some images, some assets like the skill that I have set up. People would just use the skill creator tool and then describe the skill that they want to set up. And I really believe that this is a big, big mistake in the process. What I always suggest is first of all, don't use the skill creator straight away. First of all, do the task together with Claude. And you'll see exactly what I mean by this in a second. But see if what you're trying to achieve, is it even possible? Can I walk Claude through my process step by step? And remember, you are the expert here. You want to be teaching Claude a process that you know like the back of your hand. Walk it through the process step by step. And then only after that will we turn it into a skill. So first of all, we do the task together. Next, we verify. We review and refine until the output is right. And this sort of goes hand in hand. We're going to give Claude the task, walk it through it step by step, and we're going to very importantly check: is it getting the right results? And only when it's getting the right results, then we ask it to save what it just did, turn it into a skill so that we can use it on repeat. And so let me create a skill for you live. Now what we're going to do for this demo is I'm going to see if I can teach Claude to create these social media assets. This is something that's very, very popular on Twitter and LinkedIn at the moment. These Twitter style mockups. There's a lot of thought leaders using it. We can see here Alex Homozi is using it quite a lot and getting some decent results as a way to help us post more consistently on LinkedIn. So, I want to see, can we teach Claude to create these um Twitter style mockup images so that then in the future I can just give it some text, an idea I have, and it's going to spit out these um social media assets for me. So, what I'm going to do, I'm going to save this image. And then the first thing I need to do is prepare all of the files and folders that Claude needs to get this job done. So, I want you to think: if I was going to go and actually design this Twitter style mockup in a design tool, what would I need? I would need my image that I want to swap out with Alex Hermosy's. I obviously have the original image here as reference, and then I have the font that I want to use, in this case inter, as well as the little Twitter blue verified tick. So now in co-work, let me get rid of this skill creator and let me select that working folder, and I'm going to mount this tweet format social post folder with all the relevant files that it needs. All right. And now as I said, we're not just going to go and ask it straight away to create a skill. I want to see: can I get Claude to do the task once? And in my case, getting Claude to do the task once means basically creating one image that looks like this, but has my photo there, my name, and a tweet text of my choice. So, let me ask it to do that. Can you please adapt this social media image from Alex Horoszi, which is a tweet style image format, and replace Alex's photo with my photo, which you'll find in the folder attached. Also use my name and handle, which I will provide instead of Alex Hormosis, and then replace the uh text with the following tweet. So that's my prompt. Now I just need to give it uh the name it should use and my handle, and I have a tweet text that I have prepared that I want to test it out. Now when you're giving this task, you want to give as much context to increase its chance of success. So I might add a few more things. Uh, the final image should be square 1080 pixels by 1080 pixels. I know for instance, since I've been doing it already, I'm going to ask it to please use the Python pillow library, telling it to adjust the font size so that everything fits nicely. And I'll also say the blue tick symbol is also provided. Now remember, any files that we provide here in this folder, Claude needs to actively read them and search for them. So to make it even easier, I'm also going to attach that Hoszi image as an attachment so that it gets put in the context straight away, because remember I started this prompt. Can you please adapt this social media image? So it knows exactly what image I'm talking about. So let's send that one off to co-work. And we can see co-work has already got to work, and it's mentioning that it's creating something. So let's see. Hopefully it's already creating that image. And we can see pretty nicely, after just a few seconds, we've already got the first version of our image. And I have to say, this is actually looking uh pretty, pretty good. Let me just compare this one with the uh original Hormosi image that I attached. So maybe I might say that potentially the um font weight is a little bit thin for my liking if I compare it to Alex Hormosis. And this is really the part of reviewing and refining until the output uh is correct. So, let me try and point it in the right direction. And we can actually see here it's even asking, "Let me know if you'd like the font heavier, the avatar bigger, the layout tighter, or any other tweak." So, I'm going to say, um, can you please make the text font, so the font of the actual tweet uh text body a little bit heavier to match the style of the Hozi tweet image that I shared with you. And this is really the way you want to be working as you're walking co-work through a task. Just treat it like a conversation. Obviously, make sure it has access to everything it needs. Give it the relevant information. And then you don't need to actually oneshot everything. Have it make a first draft. And now we are improving it as we go. And there we see that has updated. We can say body text is now intermedium, giving it slightly heavier weight closer to the reference. I think this is looking a lot better. And I would say it's pretty much nailed it. It has my image there, the name, everything is good. So in my case, I've done the task together now with Claude. I can see that the image has been uh created. If I open up the folder, we can see that my Yanweet PNG is there. So now the next step is simply just saving this as a skill. And we have a few different options here. So really now this is where I would reach for the skill creator uh skill. I can type this and then give it a prompt in terms of how I would want to name it skill and maybe add any additional requirements. But you can really see this is the process that Claude themselves would sort of recommend, because if we click at the top here on this um the name of this task, we can see in the dropdown turn into skill. We can turn this task that we've just done into a skill. And it simply just provides this prompt. It's that simple. And they're making it so, so simple for you to do this. Turn this task into a skill. Right? In my case, I'm just going to use skill creator and say, can you please turn this into a skill so I can create images like this on demand whenever I need them? Make sure that in the skill is baked in that the font should automatically change uh size so that it fits nicely inside the square image. And then I'm simply telling it the name that I want my skill to have. And this is important because um it's a little bit tricky to actually change the skill names after you create them. So, always make sure you lock this in straight away. In my case, I'm going to call it social create tweet image. And that's it. We can send this one off. I'm just going to close this here because what we see is that it has already now launched this skill creator skill. And so any skills that Claude is using, the skills that you're going to have it run in the future will appear here on the right hand side under skills. So, it's already used its skill creator. We can take a look just how advanced this skill creator um skill is that Anthropic have created. What's really, really nice, it also has an in-built evaluation system. So when it goes ahead and creates a skill uh in many, many cases, it's actually going to do some thorough testing of that skill, and you can actually also invoke this at any time asking it to evaluate the skill that it has created. This is all built in, and that's why I said those three layers of skills, I think it's important that you understand them: that a skill contains metadata, instructions, and resources. So, you know exactly what to look for, but you don't need to do any of this yourself. That skill creator takes care of everything. And here we can actually see, while it's creating this skill, that evaluation system working. We can see smoke testing running the script on the original tweet and on a much longer one to confirm that the auto fit shrinks. So, what it's doing now under the hood is going to actually go and try and create some tweets with shorter text, longer text, and see that the skill is doing the right job. And this is that agent loop that we talked about at the start, that it doesn't just do a task. It checks: is the work being done correctly? And now that it's confirmed that all four of its test cases worked as intended, it's going to go and package this skill. All right. And we have our skill created. It also has opened it up here on the right hand side so that we can um preview. So of course we have our skill.md. And here it lays out those step-by-step instructions of how our tweet images should be created. And we have a bunch of reference files. So we have both assets and scripts. If we take a look at the assets, it has actually saved the fonts that it needs here as well as the Twitter verified badge. And we have the Python script to actually create the image. Now this is very, very important, why again you need to understand the structure of these skills, 'cause one thing that I've noticed that it hasn't included, which I'm going to ask it about, is the fact that my image is not here packaged as an asset. So, I want to know: how is it actually going to put my photo inside these tweet images if it's not packaged in the skill? I notice that my image um is not in the assets of the skill. Um, how is it going to know to use my image um in the tweets um if this is going to be, if this skill is going to be handed off to other people. So, here I'm just going to give it that feedback and let's see what it says. And a short while later, it's gotten back to us now saying, "Hey, that's been fixed." We can see now we also have this default photo as a part um of the skill. Pretty, pretty cool. Important to know: now it has gone ahead and created this skill, but this hasn't actually been added to our cowork account yet or to our claude account. And so very important is after it creates the skill, make sure we click this save skill button, and we can see the skill has been saved. If we now either click on manage or we can go back to customize and skills, and we see always from the top the latest skills that we've created. So, we now have this social create tweet image skill. It's created the name, the description for us. Obviously, that long set of instructions in this skill markdown and the references. We didn't have to do anything manually. We just walked Claude through the process, made sure everything was correct, up to our standards, then asked it to turn it into a skill. Now, next important step: you need, of course, Claude's gone and done some evaluation, but you should always thoroughly test your skill. So, what we can simply do now is go into a brand new task, and when I type forward slash, and remember my um skill was called social create tweet images, and we can see that that is appearing there now that our skill is set up. And to actually really test this one out, what I've gone ahead and done is I actually went and created this CSV of multiple uh tweets that I want to create. I think there's about 15 different tweets here. So, I'm just going to give it this whole CSV and see if it can create an image for each row. Please create an image for each tweet in this CSV file. And we can see Co-work has successfully loaded up that skill. It's now using it. It's also read the CSV that we provided. And we can see the images are already being created just like that. Just how quick it goes. Once you create a skill, everything happens so much faster. Let's go and check some of these out. So here we have tweet number one, tweet number two. They all seem to be matching the style and format perfectly. We can see the text getting a lot bigger um if it's a smaller tweet, and a lot smaller if it's a bigger tweet. This is absolutely perfect, and you can see just how easy and quick it is. Now this was obviously quite a simple skill that we have set up, but the process is the same. You do the task with Claude once. You then go and verify that the output is correct. And only then do you save it as a skill.

All right. Now that we've created a skill, let me also show you how to update and, more importantly, share these skills with your team. Starting with updating skills. And this is just as simple as the way we created a skill. So I can either do this in a conversation. Like imagine here if all of these tweets came out horribly wrong, I could ask it to correct and update the skill. So, what I'm going to do now, let's say I want to also have this skill be able to create dark mode versions of these images. Again, I'm going to do the task first to see if it's working, and then I'll ask it to update the skill. Could you also please um create a dark mode version for all of these uh tweet images using this exact color as the background? And then I'm going to use our 9x beautiful dark navy color that we like to use and send this one off. And a few moments later, we see all 15 dark mode versions are saved. Let's take a look. And pretty cool. I don't like them as much as the ones with the white background, but it has done the exact task that we wanted. So, we've verified that it worked. So, now how do we go and actually update our skill? Well, that is simple. We just use again the skill creator tool. And this is really how it is intended, because I can also show you if I go here to uh customize and go to my skills, and then find any um of these skills on the top right. We can see here edit with Claude. So this is even the way that Anthropic is suggesting that we edit and update these skills. When I click on that one, all that it says is simply "Help me edit" and it puts in the name of the skill using skill creator. That is how you also, not only create skills, but it's also how you should update them. Now, very important is when you use this edit with Claude button. For some reason, it's taking us to the chat window. I highly recommend that process that I've just done: go to co-work, have it do the task you want it to do, have it make the changes to that skill, and once you've confirmed that it has done them, again, we just use skill creator. So, here I'm using skill creator. Please update the social create tweet image skill so that it also includes a dark mode option. Um, if the user doesn't specify whether they want light or dark mode, please make sure you ask them whether they want the image in light, dark, or both. So, I've verified that it works. I'm now asking it to update the skill. Let me send this one off. All right. And a short while later, we get the confirmation. The updated skill is packaged. We can see the steps that it took. So, it actually copied that skill file, then added the dark mode variation to the script, updated the markdown file, did a test, and then repackaged everything. Now, hopefully, as you know by now, this skill is not saved in our cowork account. Also, when you're updating skills, you need to do the extra step of clicking on this save skill button. So, I'm going to click save skill. Then, we're going to get a warning saying, "Hey, this skill already exists. I want to upload and replace," because we now have a new, updated version uh of that skill. And perfect. So that skill's been saved. Let me create a new task now and run the test. So I can just call the skill again. So here I have the text of another tweet. I want to send this one off. And what we want to see now is, is it going to ask us that question about whether we want to use dark or light mode? Let's take a look. And we can see it's already going in the right direction. And what it just used there was an ask user question tool. And this is a very powerful thing we can build into the skills we can create. We can tell it, hey, when someone uses the skill, make sure you have all of this information. And if you don't have any information, ask it. So, in this case, it's asking, should I render this as light, dark, or both? Let's test this out on out and say both. And just like that, there we have our two tweets all ready to post. Super, super cool.

Now, let's move on. Now that you've seen how easy it is to update skills, let me show you how to share them. And there's a few different methods. So, if we go into our skills section here in customize, if you are on a team or enterprise plan, so if you're using co-work with your company, it's super, super simple. You can click on share and then you have a few different options. You can simply say, should everyone in your team be able to see this skill, or do you want to share it with only specific team members? So, if I wanted to share this with my co-founder Alex, I would just type in Alex's name, or I could say, "Hey, I want this skill to be available to everyone in the 9X team." Now, once you do share a skill, important to know, some people think, okay, it's disappeared. The thing is, it's now no longer listed in your personal skills. It is now here under shared skills. And now, if someone shares a skill with you, as of right now, maybe this changes. Important thing is you do not actually get an email. You do not see it here in your platform. It is a little bit uh hidden. If you then click on this plus button and then browse skills, what you want to look for is if you click on now this shared tab, you will see any skills that your teammates have shared with you. And you can simply click the plus button to enable them. Now let's say you are not in a team plan and just using co-work uh on your own, but you do want to share your skill with um someone. The other option that you do have is if you click on these three dots, you can download the skill. And what that's going to do is going to download the skill file. It's going to have a .skill ending, but it's the exact same as a do.zip. And then you can actually share that file with anyone. They'll go into their co-work account, click create a skill, upload a skill, and upload a file. A big difference though that I want to point out between the two methods: when you just download the skill file, and then they go and import it, those are now two totally disconnected things. If you go and make updates on that skill, that's not going to immediately update the skill for whoever imported that file. They're literally just importing that current version of the skill. And this is a little bit different when you're sharing the skill amongst your organization. First of all, I am the owner um of the skill. So only I can make changes to this specific version. Of course, they can go and duplicate it and make their own, but this belongs to me. I can make changes. And now any updates I make to this are immediately enabled across the whole team. So that should have given you a good understanding of what skills are.

Let me just quickly recap about what you gain with skills, just to drive this home and why I think they are the most important feature, or the most important thing you can be working on to leveling up your AI usage. First of all, skills enable you to stop repeating yourself. You define the workflow once, you do the task together once with Claude, and then you can run it on repeat using a single command. Additionally, you're going to get faster, way more consistent results because the skill locks in the process so that every time you run it, it follows the exact same structure. I think we saw it there in those image files that we were creating. Obviously, the first time we asked Claude to do the task, it had to actually write the Python scripts, had to figure everything out. Once it actually did that, we turned it into a skill, it's purely just executing that script, and we get our images a lot faster. Additionally, we're going to use way fewer tokens because Claude already knows exactly what to do. There's no tokens wasted in figuring out the process. It just goes and reads the skill.md and executes the task. Also, we're using way fewer tokens as opposed to having all of these instructions in system prompts or project instructions. And that's because, remember, those skills are built on three layers. So, the only thing that it sees is the description and the name. It's only going to read all of the instructions at the moment it needs it. And for me, probably the biggest benefit of skills is the amount of leverage that you're going to create for you and your team. Share skills so that everyone benefits from the workflows that you've built. This is really something that we're seeing more and more. You have the subject matter experts creating the skills in their domain and basically sharing their expertise amongst the whole team.

So, by now you've seen the importance of giving Claude co-work access to your actual files and folders, connecting it to your tools, and teaching it the way you work using skills. All of these things are actually connected, and they're all part of one very important concept. And it's this concept that separates beginner co-work users from advanced ones. It's called context engineering, and that's exactly what we're going to cover in this chapter. If you've ever used chatbt or regular clawed chat, you've probably learned a few prompt engineering tricks. This is very much the old way of doing things. Things like telling the AI model that it is a senior expert in a certain topic, or that this task is so, so important that if it fails, your job depends on this and you're going to get fired. Now, these tricks were actually needed and did work when the models were smaller. But now with modern models like the ones that we are using in Claude Co-work, and with Coowwork specifically, these are not really relevant anymore. What matters now is context. Specifically, what we put into co-work's context window. I'm going to cover exactly what that is in just a second. But basically, high level, the way you need to think about it for any given task: it's our job to make sure that co-work has the right materials for the task. Does it have access to the right files that it needs to get the task done? Have we given it the right instructions on how we want the work to be done? Have we connected it to the necessary tools? And have we given the right amount of detail in the conversation? So instead of just prompt engineering, we do context engineering. And when you understand this, the results you get from co-work drastically improve. Let me walk you through exactly how it works. First of all, what context actually is. And here I'm going to be explaining the context window. Then the five levers that you can pull as a context engineer in co-work. So how can you make sure that co-work has access to the right amount of context for any given task? And then after that, we're going to jump into the live builds and live walkthroughs. We're going to put all of this together on real workflows where context engineering does the heavy lifting.

So first, let's talk about the context window and what it actually is. The way you can think about this is the context window is all the information that the model, so in this case Claude, can see when it's producing a response. So you can think of it like a box that you need to fill with information. And the size of a context window is measured in an amount of tokens. If you're using the latest Claude Opus model, at the time of recording right now, we have a 1 million token context window. And to put that in perspective, 1 million tokens is roughly 750,000 words or around 9 to 10 novels. Now, it's important to know that even when you start a brand new task from scratch in Co-work, there will already be some context used up. There is a system prompt that tells Coowwork exactly how it should work. All of that goes into the context window. Any custom instructions that you set, and I'll show you in a second exactly how to set these up, any tools that we've connected, and also any skills it has access to. And then there's obviously the part of the conversation, our first messages to co-work. Now, what's very useful to actually see this in practice in co-work, we can actually use this in-built context skill. And so here I'm doing it in a brand new task, not connected to any folder. I can fire this one off. This is a brand new task starting from scratch. And whenever we call this context skill, we get basically a summary of how much of that context window we have used up. And as I mentioned, it's in the amount of tokens. So right now in a brand new uh task, I'm already using 50,000 tokens out of my 1 million. And we get a full breakdown of where those tokens are coming from. So you can see in any new conversation, you're going to have a large amount coming from the system prompt. This is all the inbuilt instructions from the team of Anthropic telling it how to run an agent loop, how it should think through certain tasks. All of this you'll basically have when starting any new task. You'll have system prompt eating some of that context. Then we obviously have the tools that we have connected. Now the tool section that is deferred, you can skip that for now. You don't need to worry about it. Basically, they've just optimized the way that tools load so that it's not eating as much context. We can see our skills are using a certain amount of context. And then we have the messages in a conversation. Now, here, since I literally just started a new task and called the context straight away, we can see we're basically using 0% of our context window through messages. But what happens is as you start working on a task, this context window is going to fill up. Maybe we've had some longer conversation back and forth. Claude co-work has read some files. It's actually called some tools and executed some skills. And we can see if we run

That context command on a longer task that I have already had a few back and forths with co-work with. In this case, it created some skills for me. We can see how much more of that context window is being used up. Here we're almost at uh 30%. And by far this is coming from the messages back and forth. And within these messages are any skills that get executed, any files that get read, any tool calls, the answer from those tool calls. All of these go into the messages category. And this is typically what's going to eat up your context window.

So why is this concept so important? And why do you need to know about this as a co-work user or someone trying to get the most out of co-work? If we put too much uh information or too much context into this context window and as this context window fills up, some negative side effects occur.

First of all, a concept known as context rod. So, as we put more and more context into our context window, the replies from the agent get significantly worse. And I'm not sure how many of you have had that when you back in the day had a very long conversation with chat GBT and then it started forgetting things that you previously talked about. This is all because of context, right? So as it has more and more context, its ability to remember specific facts from that context gets worse and also the reasoning quality of the model drops.

Additionally to context rod, another side effect of having too much in the context window is tool confusion. As I showed you before, a part of that context window is what are all the tools that the AI model has access to? And since the AI sees every single tool definition, the more options that we give it, it's all competing for the model's attention. And if we have tools that have maybe have similar names, maybe if we are managing both tasks in notion and tasks in another tool like Asana, it's going to start getting confused about which tools it should use for a certain task.

Another side effect is that you'll also see slower responses because what happens actually is when you send a new message to co-work in an existing task, the whole context is basically sent again. We're sending the whole context every single message. And obviously then if we have bigger context means the AI needs to process more information for each message and we have to wait longer for it to respond.

And perhaps the biggest side effect that probably plague most co-work users is when co-work has to deal with a large amount of context in a task, it's going to eat up your usage a lot faster. So, if you remember at the start when we talked about the pricing, we get access to different usage limits over certain time periods depending on what plan you're on. And when you're working in conversations that have a large amount of context, it's going to use up either your uh session or your weekly limits a lot faster.

So now that you know the importance of managing your context, let's talk about how to actually do it and exactly what context engineering is. Put simply, context engineering is the principle of designing and managing the context given to an AI model so that it produces better outputs. And in co-work, the way that we can manage the context given to the AI model happens over five different levers. So these are the five levers that we can pull as context engineers. is I'm just going to introduce them and then we'll touch on each one one by one. A lot of them we've already covered.

So, first of all, starting with the conversation. This is the actual prompt that you give, the messages that you send to co-work including both the text and attachments. Then next up, we have the working folder. This is something that we have already covered. So, what files and folders does Claude have access to when it's working on a task? Lever number three is connectors. Which tools are we connecting to Claude so that Claude can pull data from them and also take actions in them? Number four, skills. What we just covered in the previous chapter, the set of instructions that Claude follows for a repeatable workflow. And then finally, the fifth lever that we can pull as a context engineer is instructions. These are saved instructions which Claude always reads first before starting any task.

So obviously we've already covered these first four items quite extensively in the course already. So I just want to take a little bit of time and talk about instructions. And as you can see, these are quite similar in a sense to skills. Both of these definitions contain the word instructions. The way you can think about it is a skill is instructions for a specific workflow, how one specific task should be done. Whereas instructions are sort of general guidelines for the way we want Claude Co-work to work. And these can be set at different levels.

So, if we head back to co-work, the first level where we can set this, if I click on my profile at the bottom and then head to settings, you'll see here in the general tab, we have instructions for Claude. And here we can give instructions that Claude will keep in mind across all chats and co-work. So, this isn't just co-work specific, whether you're using Claude chat or Claude co-work, we can add specific instructions here. So, if I actually cut out what I have set, we can see an example here. So it says keep explanations brief and to the point when learning new concepts. I find analogies particularly helpful. These sort of things is what you could put here in the general instructions. In my case, what might be useful for uh you is giving Claude a little bit of information about yourself. So what your role is, what company you're working for, as well as some things that I want to it to always follow no matter if I'm working in claw chat or co-work.

Now, we also have the ability to set instructions specifically in co-work only. In the settings, if I head to the co-work section, we see here global instructions. Instructions here apply to all co-work session. Use this for preferences, conventions, or context that Claude should always know. So, very important is whether you're adding stuff in the general instructions or in the cowwork instructions, these are always getting put into the context window and always loaded. Claude basically always reads this before it starts a task.

Now, we also have the ability to set instructions for a specific folder that we're working in. So, whenever we connect Claude to a working folder, we can give instructions that whenever any task is created inside that folder, it will follow these specific instructions. So, let me just select an existing folder that I have, which is my YouTube folder. I'll just give it a very basic prompt like uh what is in this folder just to show you how to access the instruction settings. And we can see while that's running what I'll show you on the right hand side here is next to the working folder we have this instructions and it says claude.md. And this is simply a markdown file that Claude will add inside that folder on your computer. And anything written inside this claude.md will also be read before it starts working on the task. So what where this is very useful is if you want to give some specific instructions for how it should work in that folder, how it should organize files, any specific file naming conventions that you want or anything that it shouldn't do, maybe some files that it shouldn't delete when working in that folder, you can add them here to the claude. MD. And this is easy to see when I actually open up this uh working folder. We see this markdown file. claw.md is just a markdown file that we have inside this folder. So you could also add it yourself manually, but the easiest is just to basically start a task in a folder in co-work. This instructions will appear up here. You click on it and then you can set any instructions for that folder.

Now maybe a more user friendly approach to this is something that was added quite recently which is projects. And projects are pretty much identical to a working folder. However, they just make the UI a little bit easier. So, what I'll actually head over to projects and start a new project. And here, what you can do is you can already actually grab an existing folder on your computer. Maybe you've already been working in a folder with co-work. You can start using it or you can say start from scratch. When you start from scratch, you're going to give your project a name. So, maybe I'll might say 9x live workshop prep. And so for our workshop prep, I can paste in some instructions giving it all the information it needs to know, all the context it needs to know about how it should work inside this folder and how it should help me prepare our weekly free 9x live uh workshop webinars that we run. I can also add any files I want. But in the end, this is identical to a working folder because all that happens is we select which area on our computer. In this case, it's in the clawed uh folder, clawed projects. And all it's going to do is create a folder for this project. Any files that we add here will get added to that working folder.

One additional feature that we do get with projects is um specific memory. So when you're having chats, when you're creating tasks inside this project, if you ever ask claw to remember something, so hey, don't do that again. Work this way in the future. Do this. It's going to add it here inside the memory and we will be able to see that and manage the memory for this specific project.

On the topic of memory, this is sort of something I would bunch together with instructions. Claude also does if you have memory turned on remember things about you individually. Again here we can go to our settings and if we go to capabilities we have this memory section. So here depending on your settings you'll either have memory on or off. I would highly highly recommend doing it. What basically happens and this is again something that's universal across Claude not just co-work as you start chatting as you start working Claude is going to remember certain things about you and automatically put that inside the context window. We can easily view our memory. So here it's also decided to for instance put information about uh 9X the company where I am a co-founder introducing my co-founders also what type of training we offer the typical tools that we use where our community is hosted and the price of our pro membership it also has some personal context which I might blur out here and as I said Claude fills this memory automatically. Though as a best practice. This is something that you could for instance review once per month. If you want to remove some things that you think are irrelevant, you can also absolutely edit this memory. But just know that this pretty much gets treated the same way as the instructions. This gets loaded into the context before every single task in co-work. And this is really the important thing to remember when it comes to context engineering. As a context engineer, it's basically our job to figure out which of these levers, and in general, it's going to be more than one, we should pull for a given task.

But I just want to cover before we get on to live builds and walkthroughs, some general context engineering best practices. First of all, build skills instead of pasting the same instructions over and over again. So, if there's a specific workflow you want it to follow, use a skill instead of always prompting it in the conversation or adding it in the instructions.

Start fresh chats for new tasks. And I think this is one that a lot of people are guilty with and something that often burns the usage is you start a task in co-work. It does a really really good result and then when you want to do a similar task, you just use the same conversation. I'll give you the example of creating the assets for our 9x live um webinars that I showed you earlier. Every week we run a different workshop. I've created a dedicated skill for this to come up with the images. When I want to generate the images for next week, because this is saved as a skill, it is far better if I start a new task in that same working folder in that same project and run the skill for this week's workshop because it has nothing to do with last week's workshop. That is context that it doesn't need. So, we really need to move away from using these extremely extremely long conversations that we're probably used to back from using chat GBT or regular Claude. And whenever you have a new task, even if it's related to something else, think when can I use a new fresh chat for this.

Next up, be very deliberate about what you attach and connect. If it doesn't need to read a certain file for a task, do not give it access to that file. Additionally to this, I would highly recommend using working folders instead of attachments. Remember that in coowwork, you have two ways to give co-work access to files. First of all, here in the working folder. So in this case, when I'm working in the YouTube folder, it has access to all files in that folder, but it's only going to use the files that it needs for a given task. So for instance, when I want to work on my next YouTube video, it's only going to go into that subfolder and ignore everything else. When we add files here as attachments, these get loaded into the context straight away. And so you're not even giving Claude the decision of should it use it, should it not. It will always be loaded into context. Now, where this is useful is for instance, if coork's working on some design elements and you want to tell it, hey, this is going in the wrong direction or you've made a mistake, you could take a screenshot and add it here as an attachment. That is a far better use for attachments. But in general, if you needed to crunch some data in spreadsheets or help you work on markdown files or help you work on PDFs, put them always inside the working folder.

And the final tip that I will share here is work with local files before you go connect it to live data using the tool connectors. And this is specifically important when you're working with large data sets. Maybe you want it to analyze the contacts in your CRM or analyze some data that you have in BigQuery. If we're talking about thousands, potentially hundreds of thousands of rows, it's going to be far better if you just export that data manually, have it in there as a CSV or an Excel file and then have it crunch the data. Connectors are absolutely amazing because they make that lives a lot easier. We don't need to open up the tool. It just goes and pulls the data. But they do have their limitations. A lot of connectors, for instance, will only pull in the first 100 or a first thousand rows. and then it might be giving you answers about data without seeing the full picture. So what I'd recommend is let's say you want to run some analysis on your CRM or some of your data uh analytics exports, download them as a CSV, have co-worker do the task first and then when you want to automate this and maybe have it set as a scheduled task, ask it to connect to the specific connector and ask does it see all the same data. Why this is great is because coowwork knows already all of the available fields, all of the data points it has access to. It's going to be a lot faster when connecting to the connector and it's also going to tell you things. For instance, like hey, when I was working in the CSV, there were these four fields which I don't see when I use the connector.

All right, now let's move on to the main event and four live builds and walkthroughs that I'm going to be going through in Co-work to give you the full experience because honestly the best way to learn is by seeing it in action. So I'll be doing two live builds from scratch where I'm going to be using co-work to automate something that previously I would need to be doing manually. What I really want you to focus on is don't worry too much about the specific use case that I will be building. What I want you to be doing is as I'm going through and building this out, I really want you to be thinking is how can that same pattern be applied to something that you need to automate, something that you want to optimize in your day-to-day process, because a lot of these follow very similar patterns. And I'll be calling out suggestions throughout the live builds, but I think I've picked the use cases where we sort of cover the most amount of co-work features as possible so you get a really good understanding. Now, I'll be building these live. Co-work does take some time, so I'm obviously going to cut the moments where co-work is processing, but I will not be leaving anything out. If co-work goes in the wrong direction or makes some mistakes, you will see the full process of how I correct it and how I get it back on course. After the live builds, I'll also be walking through two setups that we have for ourselves at 9x, so you also get an understanding of some more advanced setups.

All right. So, this is what we have planned for build number one. And I'm going to be setting up an email triage agent. Basically, the problem that I have is I receive a load of emails. And because I receive so many, I often miss the most important ones. I also try and not check my inbox too many times during the day to get uh distracted by the latest AI news and newsletters, which is very important that I h keep up on, but not something that I want to be jumping on every single hour. So what I want to try and do is set up my own uh assistant, my own co-orker, my own agent here in co-work that is every hour going to look at the latest emails that I have received and if there are any urgent ones that need specifically my attention or need me to act, it's going to alert me on Slack with the links to those emails, maybe some um follow-up recommendations, and basically then everything else gets triaged.

Now, the co-work features that we're going to be using during this build, obviously connectors because I want to pull from my Gmail and then send me updates on Slack. So, we're going to be using connectors. I'll most likely be setting up one or more skills for this. And obviously, because we want this to run automatically in the background, we are going to be using scheduled tasks. And I think as well, we will most likely set up a dedicated project for this. So, I'm also going to add uh in projects, but really um as I said just just before, don't worry too much about the specific use case. Where this one is good is it's obviously teaching you about using a setting up a co-work task that runs on a schedule. So, some sort of automation and really think about I think we all have this in our day-to-day where we regularly need to be checking a certain source and when something happens in that other tool, in this case it's when I'm getting emails, we need to be notified. Maybe it's when certain orders are coming in or maybe uh customer requests or tickets that need to be checked on a given schedule. It doesn't need to necessarily be every hour. It could be things that you're checking once a week or once a day. So really think about um as I'm building this out, what is a similar process where okay, I need to go and check something regularly, whether that's weekly, daily, uh hourly, and then when something happens, I need to go and take some action. And that's what we're going to be setting up um with this build.

So, the way that I'm going to get started with this one is maybe you could even just like go and take this whole spec whoops and drop this into co-work. That's not the way um that I would recommend doing this. I would really recommend doing this step by step. And so, the first thing that I'm going to check is what abilities do we actually have in these two connectors. So, let's head over to co-work and check out these two connectors to start things off. In my co-work, I'm going to head to customize and then connectors. So let's first check with Gmail where we're going to get the data from. And as I um mentioned before in the connectors uh part of the course, always important to quite regularly refresh the tools list. I hope this is something that starts becoming uh automatic. But what I have noticed is that these connectors are um often being updated and improved and you actively need to update these tool lists to make sure you're working with the latest version. All right, so let's have a look at our Gmail. What we can obviously do is we have a bunch of readonly tools. So we can um retrieve a specific email thread and we can also search email threads. So this means it should be pretty good to pull in the latest emails. That's probably the first point I will try. And now what I want to just see is how I can actually go about triaging the email. So we can create new drafts, list drafts. Here we go. Okay, nice. So we can add labels to a thread, remove labels from a thread, and add labels to a message, remove labels, and we can also create new labels. Very, very nice. So I think that is actually going to be the method that I'll be using. So what I needed to try and figure out is because this is going to run on a schedule, I want to make sure that we're not checking the same emails twice. So, if an email has already been triaged by the agent, the agent's already made a decision on is this important to me, yes or no, I don't want it to check that same email again in the next hour. Now, obviously, we could put in some sort of like time filter, say, hey, only pull in emails in the last hour. But what's important to know is that um co-work when we set up the scheduled task later on, this will only run when co-work is open on your laptop. So, what I'm thinking here is I'm actually going to get it whenever it triages an email, it's going to add a certain label, and then I'm just going to get it to search for new emails that have come in that do not have that label attached. So, we know that they haven't been processed yet.

All right. So, now let's go and check out our um Slack connector. So, here we have interactive tools, readonly tools, and write tools. Okay, cool. We don't need to really read only. So, this is fine. Basically, what's nice is here. Here we have everything we need. We can schedule a message and send a message. Yeah. And right now this is on needs approval. Um but that's totally fine. So I think we have everything we need in the connectors. And now we can actually move on to um building out this uh agent step by step. So what I'm going to do is I'll go back to co-work. And now here I could kick off a new task. In this case, as I said, what I'm actually going to do is create this as a project. I think it's quite nice whenever you're setting anything up on a schedule. um projects give us a little bit more flexibility also for u maintaining memory and these sort of things. So it's exactly the same as if I would have it now working in a dedicated folder. In this case, this email triage agent doesn't really need to have a local folder, but h let's see how we go and like eventually build up on this. Maybe we're going to be getting it to like download attachments from emails or all these sort of things. So I recommend whenever you start something out, we're always going to give it a folder. In this case, that will be a project. So here I'm going to go to my projects, create a new project, and this one I'm going to start from scratch because I don't have an existing file. I'm going to call it email triage agent. And I'm not going to give it any instructions or files yet because we're going to uh do that as we go. So let's go and create this one. And as we mentioned before when I touched on the uh projects in the context engineering part of the course, all this is doing is literally just creating a folder for you on your local machine. So let's go ahead and create this. And now we can start working through this task. Now what I recommend is not just basically bombarding it with the whole thing that you um want it to do. that will probably work in maybe 50% of the cases, but I really like the approach of when you set something like this up is working through it step by step. So, the very first thing I'm going to ask, can you please pull in my latest five emails um from my Gmail? So, a super super simple task just to see if it's able to do a very key part of the workflow.

So, we've got here the last five emails um from my Gmail and we see that one of the emails it's pulling in is from Twilio. But when I actually go and check my emails here, I noticed that Twilio is actually in my um promotions part of my Gmail. So, I'm going to see if there's a way that I can get it to only pull in. I just wanted to basically check the last five emails here in my primary uh inbox. And this is where the benefit is of doing this step by step. If you just give it the whole task, it'll be very hard to actually find these little quirks of maybe where it's gone in the wrong direction, and you have a lot of debugging to do. So that's sort of your two approaches is dump everything on it and then you have a lot of debugging or you work through it step by step and basically hone this uh along the way. It looks like that is pulling in both emails from my primary and my promotions tab. I want to make sure that it's only pulling in emails from the primary tab. Is this something you're able to do? So now I'm just going to give co-work that feedback telling it that right now it's pulling in not from not only from my primary tab. I'm even going to say share a screenshot um of my inbox. So to give it some more context. All right. And now we get the answer back from co-work. And you can see it's added in this additional query here. So it has category primary in inbox and the page size is five. Our last five emails. And now this is what we have here. Now the next interesting thing I notice here is when I'm pulling in these last five emails, these are just giving me the last five emails. let's say that are the first in the thread that I have received. If we see here, it's actually pulling in this one, which is the sixth email superhuman. And ignoring this one, which is an email thread, there's already been 11 messages back and forth where um we're working on a new webinar collaboration with a team at uh customer.io. And again, now all I need to do is give co-work uh this feedback. So, I'm just going to take a screenshot of this and I'll point out to this specific uh email. And this is a really good way of just basically giving um co-work the feedback that it needs with basically just providing uh screenshots of the real data. So I see here you're missing the um email thread that I'm having with customer.io. Could it be that you're filtering based on the first email in the thread and not the latest emails that I'm receiving? if that makes sense. And this is really the way that I would just approach this is just treating it like a conversation, having this back and forth. Yes, it does take a little bit longer, but you're working through it with co-work, teaching it the way you want it to do things until you are happy with the result.

All right. And so I've done some debugging here together with co-work and it actually found a little bit of a quirk with the Gmail search tool. So interesting when it uses the search threads tool, which is actually the only way that it can search. So it can only search threads. It can't search specific messages. If I basically ask it to pull in the last five threads, it only does this based on the first message in that thread. So that's why that email with customer.io that has basically a few back and forth replies is not coming up. So we'll need to find a different workaround. And what the agent suggests, the workaround that we basically found is instead of only looking at the latest five uh emails, what we're going to use is additionally to the new labels that we create is also filtering only unread emails. So the way this one's going to work is I'll basically get the agent to pull in uh any unread threads. And what's great is after it pulls in the unread threads, it's going to mark them as red, which it is able to do. And then if a reply comes in that we again marked as unread and should again meet the filter. And here we can see it was able to pull in those five emails that we wanted it to pull in. And now we can basically move on to the next step.

So what I've basically done is in my Gmail I've set up a few different labels that we're going to have our agent use. Any emails that require my attention. I want it to label as action required. Any emails or email threads that I need to be informed about. I want it to tag as FYI and everything should basically get tagged as triage. So I know that the agent has processed it and the agent is also going to mark them as red. So to prompt my agent with this, what I'm going to do is just take a screenshot of these labels. I'll copy that over. All right. Now, basically, I've spoken out this prompt. I've just cut forward so you don't need to to listen to me sort of ramble my thoughts. But what I've basically done is um attached the screenshot of those three labels that I want to use. and then underneath each label told it a little bit the conditions that when it should apply each label. Now, this is only very much a first run. We don't need to nail it straight away. The very important thing is we're going to test and iterate on this. So, for those last five emails, I want to see if it's able to correctly triage those. And now, because we've gotten to a point where we're not only reading from Gmail, but we actually want to take some action. So, now Claude wants to apply a label. And here, it's asking me for permission. What I'm going to do is I'm going to allow for this task and we'll see later on how we can um automate this so that we don't have to like accept this every single time.

All right, very nice. So it's now come back and um basically had the first run and I think it has actually nailed it. So let's have a look. So all five labels applied to require action. So, one from our video editor, uh, Thomas, who's basically sharing a new edit from one of our YouTube videos, and we have the custom row webinar collaboration that needs an answer. So, I've got a question from Giorgio that I need to get back to. And then we have three triaged emails. Very important. Now that it's actually done some work, we need to check if this is correct. So, I'm going to head over to Gmail and refresh the page. And we can see those labels are applied. Very, very nice. And for instance, if I take a look here at this one that it mentioned as triaged. So even though this was addressed to me directly, we can see this is sort of like a cold uh email selling some sort of service. You can see here that there is an unsubscribe button. So this absolutely correct that it marked it as uh triaged. And as well the other one was an N8N security update. I don't need to take uh direct action. This is just N8N telling us that there's some changes in their platform. And also we have here a newsletter which also got marked as triaged. Now what I'm going to do is I'm actually going to mark these all as unread again. And let's finish off the process with Claude. So what I'm trying to do here is I'm getting Claude to do the whole process over a small batch of emails and then we're going to turn it into a skill and schedule the task.

Perfect. Now that you've processed all of these five emails, could you please mark all of them as red? So basically removing the unread label again. We need to allow Claude to actually take this action. So I'm going to do that. And it's confirmed that all five threads are now marked as red. Let's go and take a look. Let me just refresh the page. And there we can see that is correct. And everything else has the label applied. Looking very good. Now remember when we go back to what we want our agent to actually do, anything that it sees as important. So in that case, the email from our video editor Thomas and the customer o thread, we want it to alert us on Slack. But before I go and do that, what I'm actually going to do is I'm going to save what we've got so far as a skill. So here I'm just going to call the skill creator skill. Can you please create a skill called triage emails, which basically saves everything that we have just done. So, it needs to look for any email threads that are unread and that do not have the label triaged. Confirm whether you think that's the right approach. That makes sense. But basically, then whenever this skill runs, it should go and check the latest threads and mark only those with action required and FYI, and everything else gets marked as triaged. So, whenever you're using the skill creator, obviously what's great is it has context over this full conversation. So, it knows all of the decisions we've made up until this point. All I'd recommend is telling it the name of the skill that you want it to create and some additional context that I've given it. So, I'll send this one off.

Now, Co-work has gone ahead and started creating this skill. But because I actually asked it its opinion, so asking it if it thinks it's the right approach, what it's doing now, it's asking me some follow-up questions. So, here it's asking me what search query should the skill use to find threads to triage. Now, this is a little bit uh let's say cryptic and so if you ever aren't sure about something it's saying, you can always ask it a question. So, you can type a question here. So, in my case, what how I determine this is here it's basically saying that as soon as a thread is classified, it should never alert me again. I don't actually think that's the right approach because imagine is uh the first email in the thread I get alerted from Slack. I go and respond and then they reply to me again and I need to act again then I want to be notified. So here I'm going to say I want it to reprocess um action required and FYI labels. Uh I think it could have been a little bit clearer here in its writing when there is a new reply. Should the skill mark threads as red after labeling them? Yes, absolutely. That's important. Otherwise, this whole thing doesn't work. How should the skill run? Uh, in this case, I'm going to set it to on demand because scheduled is something that happens later. So, and this is a little bit of a misconception is even if the skill says something like run every morning at 8:00 a.m., it is not going to happen. It only happens when you use a scheduled task. So, this doesn't really make any sense that they're suggesting it. I'm going to say on demand only, but then I'm going to set up a scheduled task that triggers this skill.

All right. And after just a couple of minutes after answering those questions, Co-work's now finished and it's actually created the skill for me. It's given me some feedback. So the skill is built and packaged. I just need to click on the skill file. It says above, but it's actually uh below here. So what does it do? It looks up the uh IDs of the three different labels. So it knows how to use them. Meaning if I change the label name, it will be able to adapt accordingly. Then it searches my inbox. So here it says if the message is in the inbox is not one of the promotion or social categories but it is uh unread and does not have the label triaged. Interesting. Here it says that the page size is 25. That should be fine if this is checking every single hour but already something to keep uh an eye on. For each thread we get the full content. So we're not just looking at the subject lines of the different messages in the thread but getting basically all of the information to make a decision. And then it does its go and does its classification. Now it basically suggests a couple of things to do once we've installed this is we should test this on a fresh batch. I think that's very good. Um eventually schedule this and then iterate on the rules. So some really really good advice. Um we can take a look fully at this skill that it has set up here. If we take a look at the skill MD, this is a very very long instructions. You can definitely go through and read through this if you want. If we notice that we start testing this and some things are wrong, that's when I would go back and take a look at why some of those decisions have been made. This is the skill MD. But if we actually look at the skill file um that has been created, we see that actually it is just a skill.md. It doesn't have any additional files, which makes uh sense. It doesn't have any reference files that it needs. Now, in order to actually set this up and import it into my co-work, I need to click here on save skill. And now that the skill is saved, I can click on managed. And we see here this triage emails is here inside my personal skills. And at any time we can go and edit this just again using the skill creator to update it like I showed you in that part of the course. So with this skill now saved, I can iterate and move on to the next step which would be um sending the Slack message to the right channel.

So what I'm going to do is I'm going to go into a new task in that project. So I can click here just new task. I'll make sure my email triage agent project is selected. And now what I actually need to do, I haven't had any new emails come in. So what I might do is I might mark this one as unread. So this is basically recreating the experience. Imagine if this got marked as action required when the first messages in the thread came in. Now there's been a new reply. Is it going to pick that one up? And I'll also select a couple of emails that we haven't triaged yet. I'll also mark them as unread. So they basically are getting searched as well. And now in co-work all I can do is basically call that new skill. And let's see what the output is. And we can see it's called that skill. It's now given itself the to-do list. It's connected to the Gmail connector uh already. And it's working through its tasks. All right. And that happened a lot quicker because now it knows exactly what to do. And we get our result. And we can see it successfully triaged uh two of the newsletter emails and that um thread that I marked as unread did get um alerted here which is perfect. So now I can move on to the step of sending this to Slack.

So what I've gone ahead and done in Slack, I've created a new channel and this is where the agent is going to alert me of any emails that I need to take action on. So the first thing let's go on now and take a look at the summary um that we received from our triage email uh agent. So now what I want to do is I'm going to give it some instructions to send me a message on Slack alerting me only to the action required and FYI email. So in this case we should only be alerted to this email for the webinar collaboration. And I've asked it to please include a link directly to the email so that in Slack I can easily just open it and then of course reply because in most cases action is required from my side. Now, very important is I need to tell it where it should send the Slack message because when we use the Slack connector, we're giving it the ability to send messages in any channel that we are in. And we obviously don't want it to accidentally send this to one of our colleagues, one of our co-workers in a public channel. So, the best way to do this is really lock this down by the channel ID. So, that channel that I've created, if I click on the top here, I get the channel ID. So, I can just copy this one. And I'm also going to give it the name of the channel. So I'll paste in here the channel ID and I'll also grab the name of the channel so that it is not confused. So let's go and send this one off. And it's asking me for approval to send the message. We can actually see that draft message here that it's sending. I'm going to say allow for this task. And that is set. Let's take a look at the message. Perfect. Then we have inbox triage. One thread needs your attention. We have it listed here. And in one click, I can click on that one. And let's open the link. And here we might need to blur some of this out. But we land exactly on that thread. Very, very cool. This is looking great. Now, if this is the moment where you actually say, hey, I want to adjust the formatting of this. I want actually the whole email to be printed here. I want more information. I want to know times and dates. This is where you would just go back and forth and update that. But I want to get this first build out and shipped and moving along. So what I'm going to do now is I'm going to actually go back to co-work and ask for its opinion. I would like to set this up as a scheduled task so that it always uses first the triage email skill. Once it uses the triage email skill, if there are some emails that are action required or FYI, it would then go and um send the Slack message. Now, I need to make sure that it always sends the Slack message to the exact same channel. Do you recommend me creating a separate skill for sending the Slack message with all of the details of the format of the message, which channel it should get sent to, and then having that in the scheduled task, the instructions to run both, or should I incorporate everything into the one skill? And this is something that could happen many times where you're not sure of the best approach. And I really recommend just asking co-work here. It is super super intelligent. It knows its own system. Obviously, sometimes you need to push back a little bit. But here I'm basically asking it does it think that I should create a separate skill now. So I have one skill for filtering the emails which I could use potentially in other workflows and then I have another skill for alerting me on Slack or should it be one skill that combines both? And here we have our answer and this is something that I was predicting as well. So it actually recommends us creating separate skills and it says for three reasons. Single responsibility triage emails is already clean focus skills that classifies and labels. The channel ID and format live in one canonical place. If you ever want to change the channel, tweak the formatting. This lives in a separate skill and both are basically reusable uh on their own. So here's the recommended setup to create a new skill. Something

like alert triage results on Slack. And then what we're going to have in our scheduled task that we set up in a second is going to be a simple prompt saying run triage emails then alert on Slack.

Now it asked for one clarification um about how the email should be triaged. I just basically said that the triage skill already uh takes care of it and it says yes. I got it. Makes sense. And now we can proceed with creating our second skill. So I'm just going to say yes. Build the skill using skill creator. And now here we have our second skill package. We can see here alert triage results on Slack. Let me also save this one. And with that saved, now we can head over to actually schedule this task. This is going to be basically running every single hour.

We have a couple of different options here. So at the top of any task that you're working on, you can see that you can turn this into a skill or you can schedule the task. Now, all these do, if I click turn it into a skill, it literally just drops in the prompt, turn this task into a skill, and if I select schedule, there's also a builtin skill from Claude Co-work that helps you set up the scheduled task. In my case, I just want to show you how you would go about doing this manually. So, I'm going to head over to scheduled and then create a new task.

When we create a new task, we have two different options. we can create with Claude. And that one just actually executes that same schedule skill where it's going to basically ask you a bunch of questions, you answer, and then it will set it up for you. In my case, let me just show you how you do this manually. So, I'm going just going to say hourly email triage. And the description is just going to be uh check my email inbox every hour and alert me on Slack to important emails that need my attention. And now we give the actual instructions, the prompt that that will basically kick off that will run every time we run this scheduled task.

Now what I need to do is I need to make sure I'm going to be working in that same project in my email triage agent uh project. The model I'm going to select as we can either leave it as default model which means as the models improve it's just going to be basically bumped up to the next model or if you want to lock in a specific model. And now here I'm just going to drop in that very specific prompt. So, I'm going to tell it to run the first skill. So, now I can go to my skills and run triage emails. And then if there are any emails marked as action required or FYI, alert me on Slack using. And then I'll pull in the second skill that we created, the alert triage results on Slack. So, we're doing the task once we've created the skills. We've saved those skills and now literally all this scheduled task is doing is calling those skills into action and we can provide any additional context here.

Now when it comes to frequency, I'm going to run this hourly and it will run every hour. Scheduled tasks use a randomized delay of several minutes for server performance. So it's not going to run at the exact same minute every single hour, but roughly once an hour this will run. Now, I think I mentioned it earlier, but just to reiterate here, co-work needs to be open for this to run, which means that if a scheduled task is due to run and co-work isn't open, then the next moment that you open your laptop, it will basically run that previous task. So, that's why it's very important in the way that you create your scheduled tasks here, like the way I've done, I haven't said look at all emails in the last hour, because imagine overnight I close my laptop for 12 hours. Maybe I get some important emails coming in overnight. If I only tell co-work to check emails the last hour, all of them will basically be missed. So here I've created a way where it's only checking the unread emails that have not been categorized yet. Means it should never miss any important emails. So let's save this one.

And now of course what we can do is we can sit back and wait for the first time it schedules. But me, I love to be more hands-on and actually test this out. So, we have here in the three dots, we can run this one now. And perfect. We have actually had another email come in, which looks to be another newsletter that we want to triage. Let me also go and update a couple of other emails. This email from Thomas, our video editor. I'm going to mark that one as unread and also remove the label. So, it should basically treat it as if this email has only just come in and hasn't been viewed yet. And so, we should have one here that is triaged and one here that is alerted. So now with our test data ready, I can just click run now.

And we can see here that whenever the scheduled task run, it is just like you creating a task yourself in co-work. So this is just a chat where the schedule task run and all it's doing is literally sending that first prompt. As we saw, Claude always the way we've got it set up is whenever it takes action in our Gmail, Claude always needs permissions. Now, this is a bit obviously counterproductive when we're doing scheduled tasks and we just want it to do the thing we want it to do. So, what's great now is when this is a scheduled run, we have the option to allow this action for all scheduled runs. I know that the way that I've set this up with these very specific skills and these very specific instructions, it's not going to really mess up adding labels to my email. I'm going to allow this for all scheduled runs. And now, it just has the ability to do this whenever it wants to. It also needs the ability to remove emails to mark something as unread. I'm going to allow for all scheduled runs. And so our triage is complete. It's now running the alert triage results on Slack. Again, it's going to ask us for permissions. I'm going to allow for all scheduled runs because I know it's only going to be sending to this specific channel because I've locked that in in the skill. And here we can see that message has landed in Slack. Absolutely amazing.

As I said, formatting we can take care of afterwards. What's the best behavior now is if for instance we have this scheduled and it starts doing things that you do not want it to do. What's great is you can go back into co-work find the exact scheduled run where it made a mistake. So if I go back to scheduled I now can click into this hourly email triage and we have our history. So every hour it's going to run and I'll be able to click into those chats here where it made a mistake. We can then basically correct that mistake and simply use our skill creator to update the necessary skills or if we ever need to, we need to go and actually edit the initial prompt here. We can simply edit and adjust this prompt and correct it. This really is an iterative process. You're not going to nail this the first time. The idea is get something up and running quickly, have it run on some test data, and then iterate as you go. Now, from within this scheduled task view, we can also see all of those actions that we've already allowed, which is super super nice. And if you ever want to revoke any of them, you can simply remove them from here. And again, we can trigger the run manually whenever we want to do some more tests. So, that's it with build one done. We connected to Gmail and Slack. We created a couple of skills inside a project and put it all together in a scheduled task. As I said, please now really think about how you can use a similar setup for a use case that you have. Maybe if you always need to go and once a week download certain files from Google Drive and do something with them. Or maybe you need to go and check a certain notion database and adjust some settings on a given schedule. Or you need to go and manually process some orders that come in on a Google sheet. Whatever it is, check does co-work have the connectors. Do the task with co-work together once. Set up the necessary skills and then schedule. But most importantly, this is an iterative process. So build it step by step. test, run it, and improve as you go.

All right, now let's get on to the next live build. And if you remember back from the earlier chapter where we spoke about connectors, and you would have seen this slide where I said that there's three different ways to interact with tools in Claude Co-work. First, we have the inbuilt connectors, which we've already covered. But then what happens if a tool does not have an in-built connector or a custom MCP server? Another way that Claude can interact with tools is by Claude Co-work actually writing scripts and running those scripts on its virtual machine. So it can connect via API or CLI and then we also have browser automation. And what we're going to be focusing on in this live build is having Cowork write some code and connecting to an API. Now one call out I do need to make this is a a little bit more advanced. So, we're really going from the beginner to the advanced live build. And also, you will need to be an admin or have a good relationship with your Claude Co-work admin. Um, because it will require enabling and configuring a few settings in Co-work.

All right. So, this is what we will be building for build number two. I'm going to try to connect co-work to the YouTube API so that we can basically pull in all of my YouTube content, the thumbnail, the descriptions, the transcripts, even the video files, everything into a structured folder so that we can have AI both analyze it and help me improve in how I'm delivering my YouTube videos or planning maybe what YouTube videos I should work on next as well as being able to repurpose the content in other social media. Now, one thing that um I just pointed out obviously is there is no in-built YouTube connector or MCP server. If I search here in co-work in the connectors and browse connectors and look for YouTube, you can see here no connectors match your search. Or even if I look for maybe the Google API, we have Google Drive, we have Google BigQuery, Google Calendar, but nothing for connecting with the Google API and connecting to YouTube. So that's why for this build we are going to be getting co-work to run some code on its virtual machine. And it's going to be able to do this by connecting to the YouTube API and running everything locally. So we're going to be providing the our YouTube API key locally and then also configuring some network settings in claude co-work. Now the important thing here is not this specific use case around the YouTube data. This can be applied to any pattern for any tool that works with an API. So, if you have API access to a tool, you don't have a connector, you don't have an MCP, this could be a route that you can go, obviously depending on the admin settings of your company. But I'll be showing you all of these along the way. So, let's get into it.

Now, the way I'm going to start this build is first of all by teaching Claude about the API that we want to connect with. And in this case, that's obviously the YouTube API. So, what I always recommend is going and grabbing the API documentation. And I'm going to share this one with Claude. So here we have the YouTube data API reference. We can see we have a bunch of actions on the left. We can grab our thumbnails. We can grab the videos. We can grab the cap captions. These are the transcripts of our video. So this is everything that we want so that we'll get Claude to download this from all of our videos so that we can analyze them, but also so that we can reuse some of that content in other social media assets. Now, what's very nice that I'm noticing more and more in API documentation is I have this whole this option to basically copy all of this as markdown. So, that's what I'm going to do. And then I can head over to co-work. I'm going to be working in a new task here. So, first let me select folder. I'll go and create a new folder. This time I'm just going to be using a folder uh not a project so you also see what that looks like. I'll just drop this one on my desktop. New folder and I'll do 9x YouTube uh audit and create. And that's the folder we're going to be working in. I'm going to say always allow. I'd like you to connect with the YouTube API um using scripts and code that you run. Here is the YouTube API documentation. And now very nice trick. I'm going to do three back ticks. And this works in the clawed app. And then if I do space, it opens up this little code block. It's going to paste in all of that API documentation markdown. I'm also going to give co-work access to the link so it knows exactly where this is. And now what I'm going to ask it to do is actually try and get some data for one specific video. Of course, I could ask it straight away to go and like crawl the whole channel. But as I said earlier, and I think I've repeated myself throughout this course more and more, is try and get co-work to do the task once. Once you see that it works, then go and have it run on a larger batch. If I go and get it now to run on the I 60 100 videos that we have on our channel, if it makes a wrong decision and runs that across 60 videos, you're wasting a lot of time. So, I'm just going to get it to do it for one video first. So, what I'm going to do is on my channel, I'm actually going to grab the video link from a video I did a month ago on eight co-work use cases. Um, also feel free to check this one out. I will leave a link in the description below if you want to see eight Claude Co-work skills that I've set up that I use every single day in my business and you can actually download all of those as well. But I'm going to just copy this link.

So now here I'm just giving it the details. So for one specific video I want it to download the transcript, the thumbnail, and then I also want it to create a markdown file with all of the information about the video, what I call the metadata. So, the title of the video, the URL, the description, the publish date, and then any statistics that we currently have, and I'll say put all of these files in a nested subfolder, which is the video's name prefixed by the published date. All right, so that is my long set of instructions to co-work. Let's go and send this one off.

All right, so Co-work started working on this task, but it's already come back with its first question. So, it says, "The official YouTube API can't download captions for videos that you do not own. How should I get the transcript?" Now, that's a very important point because I probably didn't give enough context here in this first message. I didn't tell it that the YouTube video link that I shared with it. Let me just scroll down and grab that one. Uh, here we go. So, this YouTube video, I didn't tell it that it was a video on our channel. So, I'm just going to correct it. So, whenever it asks you a question, if none of these answers match, I can just basically also write something here. So, I'm just going to say, "Sorry, I forgot to mention, we're only going to be using the YouTube API to get data about videos that are on our own channel." So, videos that we own. All right.

So, now it's asking, "What credentials can you give me access to?" It says, "This determines whether I can use captions.download or need a fallback for the transcript." So, it's asking API key only, oorthth credentials only. Not sure yet. I'm actually going to ask it because I guess it's found something out here. Um, so I'm going to ask it which credentials do you need so that you can download the captions. Tell me which ones you need and I can make sure I can provide them. So, in cases like this where you're also not really sure, always feel free to push back and ask co-work follow-up questions um if you're not sure of what you need. Okay, so it says to download captions through the YouTube API, I need oorthth 2. An API key alone won't work. Here's exactly what to provide. And so it's giving me some instructions that I need to download a client secret uh JSON as well as a refresh token. And of course, if you're not really a technical user, a lot of this might not really make sense. But um the way I would just basically treat this is anything you're unsure about, just ask and it's going to walk you through all of this step by step. I actually already have my uh client secret JSON. I can show you exactly where we grab this one in a second. And what I like here, it says, "If you'd rather not generate the refresh token yourself, just give me the client ID and client secret." And I'll run a one-time consent flow. And additionally, it says optional but useful, a YouTube data API key. So, I'm going to be providing this one as well. And now, at the end, it's saying send me the client secret JSON. What I recommend is I'm not actually going to upload that here, but rather I will attach it to the working folder and also so that I can provide co-work with the right API key. What I'm going to do is the following. Can you please create av file in your working folder, I will add the YouTube data API key in that file and I'll also drop in the client secret JSON in the working folder as well.

And so now I just want to quickly talk about this. env file. This is a specific file type that developers use when handling sensitive keys like API keys or API tokens. And so if I open up this working folder here, I'm just going to click go to folder. So we have now the YouTube audit. When I click into this folder, we see here this is still completely empty. That's because those files are actually hidden files. And I'm not sure what this is on Windows. You'd have to look it up. But on Mac, if I press command, shift and then full stop. You see here this env file appear. So this file is here, but it is just hidden. It is a hidden file that is not normally in view. Now what I can simply do is open this up with a text editor. I would also suggest um using tools like Visual Studio Code or any sort of code editor that works really well with these environment files. I'm just going to open this up with text editor. And we see here it actually has Coowworks filled in some information for us. So it says here I should basically put in my YouTube data API key. And then here we have it's also asking for the client secret. I'm going to put that in a different location. But now all I'm going to go ahead and do is paste in my YouTube API key here. And I'm going to obviously do that off camera so that I'm not exposing it to everyone. And to show you how I would do this, for instance, if I didn't know where to download these things from, I'm just going to ask it uh can you please give me the instructions on how I can get the uh oorthth client secret JSON and then co-work basically gives me the step-by-step instructions that I need to do. So, I need to head to this console.cloud.google.com. I need to enable the YouTube data API, configure the oorthth screen. So, it really lays this out in uh perfect detail. I'm not going to waste your time with all of this, but that's simply the way you would approach it. When you want it to connect with an API, give it the API documentation. Ask it, hey, how can I connect with this API? And it's going to give you all of the instructions. So, now followed those steps. And inside my working folder, I have this uh JSON file that Google needs to authenticate. And I've also updated this environment file. Now, Google is probably one of the most complicated services to connect to. Don't worry, in most cases, you'll probably just need to create this environment file and drop in an API key. But this is probably an important time to point out that while this env file, this is the current sort of best practice workaround that most people are doing when they want co-work to connect with external APIs, but you really should check your own company's internal uh tech policy around storing API keys locally, allowing a large language model like Claude, giving it access to API keys. This is something that sort of different companies have uh different processes when it comes to that.

So now that I've given Claude everything it needs, I'm going to say good to go. API key and JSON file added. So now that Claude has everything that it needs or everything that it thinks it needs, it's put together its to-do list. So it's going to pull the video metadata. It's going to download the thumbnail, authorize oorthth and download transcript, create a metadata markdown file, and organize into data subfolders. And remember all of this for just one video and then I'm going to ask it to run through my whole system. Now we already have the very first uh issue that has come up. We get a 500 internal server error. So this is a serverside issue usually temporary. Try again. It could be that this is just an issue with uh claude. Let me actually try this again. So sometimes uh basically like what you need to know is that the claude desktop app is actually just connecting with the clawed server via API and sometimes when there are some issues you might come up into these status errors. The one thing that I would always recommend checking when this happens is just going on Google and searching for clawed anthropic status and you here you can see the clawed status page and you can see here sometimes it does look a little bit like Christmas. Red basically means it's having some issues. Greens means it's fully operational. So here right now it says all systems operational. That must have just been a small glitch. But this is going to be the place where you see for instance all of Claude is down because of some unknown reason. They will report it here.

And now we hit our first hurdle. And this is something that I really wanted to to showcase here. So I didn't just set this up behind the scenes. What it's saying is if we take a look here, network check shows the sandbox proxy blocks all Google domains. Google Aapis.com as well as the YouTube image.com. All of these are refused. This sounds very very technical. What does it mean? Well, as I mentioned to you, the real main difference between claude coowwork and claude code is that co-work runs in its own virtual machine in something that's known as a sandbox. And what that means is inside that sandbox, it's not allowed to just go and make API calls to different domains. So if we take a look at the uh YouTube API docs, we can see in the API documentation, every API call that interacts with the YouTube API actually happens with this specific URL. So this what's known as an endpoint. So it's www.googleapis.com and then it has/youtube/v3. Right now the sandbox blocks every single domain except the ones that we actively allow. So, what I'm going to do now is I'm actually going to grab this www.googleapis.com. And for this to happen, I need to go into co-work and it's actually the organization settings that I want to head to because I'm on a team or enterprise plan. And then within the capabilities section, what we have here is this code execution. Now, this is something that will need to be enabled by your admin, we see here code execution. And then we also have this section allow network egress. Here we have this domain allow list. So whenever we want claude to interact with a certain website maybe to be able to download images. This is not to do with a normal web search. But if Claude's going to run some code and download some images or interact with an API, we have to specifically allow it here in the domain allow list. And this is something that only a Claude admin can do. So what I'm going to do here is in my additional allowed domains I can just paste in the domain of that API that I'm using in this case Google APIs. Now what I can do is I can specify the exact subdomain that I want to use or I can also use this star symbol which means any uh network calls to this domain whether it's www or API or whatever it is if you trust the whole domain you can also add it this way. Now, the one thing that is a little bit frustrating, even though I've now added this domain to my allowed list, this allowed list is only loaded when you start a task from the very first moment. So, if I tell it now, hey, I've and I can actually do that. I've added Google APIs to my allow list. And so now it's going to say that it's going to retry the sandbox connection, but the sandbox is still blocked. And it's annoying that Claude doesn't actually realize this. What needs to happen is we need to start a brand new task. So, what I'm going to tell co-work turns out that um whenever I add a new allow list uh domain, we actually need to start a brand new task so that it can take effect. Can you please give me a prompt that I can use to kick things off so that you know exactly what you are working on and where you're at? And we can see co-work here saying here's a copy paste prompt for a fresh session. And I just need to copy this. This is that one little a little annoying piece, but it's not too bad. I'm just going to start a new task. I want to make sure I'm working inside that same working folder that I have set up because this working folder has my API key, has the Google authentication JSON that uh we downloaded. And now I can basically paste in this prompt to kick things off and we should be good to go.

All right, now comes the point that we actually need to do the oorthth. So oorthth basically means similar to like using the inbuilt connectors. Whenever we connect with a connector, it's using oorthth and we basically just need to log in on our behalf to give co-work the access that it needs. Now, in this case, because we're doing this work, it is a little bit more cumbersome. What I need to do is basically copy this link that it has given me and then it's asking me after I approve, it's going to try and load a certain page which has a code. Just copy the whole code and paste it here. So, I'm just going to go to my browser and paste in that URL. Now, I need to connect with the account that runs our YouTube channel. So, I'm going to do that. Now, I need to give permissions to see, edit, and permanently delete your YouTube videos, ratings, comments, and captions. I'm going to allow. And now, we get to this page that looks like, hey, it hasn't worked. But actually, this URL contains everything that Claude needs. So, I'm just going to copy this one and paste it in here. And amazing. Here we have all of the files that we need. Just like that. This one actually came together pretty pretty quickly. We have, of course, our beautiful uh thumbnail. Here we have, of course, the full transcript of what I say uh in the video, both in text format, but also in the timestamp format, which is going to be super nice, for instance, for coming up with like the chapter, thumbnails, and uh descriptions. And then we also have this metadata document. So you can see here the title, the URL, the duration, how long it is, what date it was published, the current statistics in terms of views, likes, uh comments. Sadly, no one favored that video. So maybe if you want this video to be better than that one, feel free to uh favorite it. That would be very, very nice. And that we have the full um description. So this is amazing. All of these files. Let's actually go and take a look. I'm going to open up that working uh folder. And we can see it did exactly as instructed. It created its very own subfolder with all of that information in there.

Now, of course, what I could do now is just go and ask it to do this for all the YouTube videos in my channel. However, remember what I've been trying to teach throughout this course is when you get it to do something good that you want to do again and again, this is a great time to turn it into a skill. So here I'm going to say use the skill creator. Can you please create a skill for getting all of this data for a specific uh YouTube video? So I would just provide the YouTube video URL or maybe a channel ID for all the uh videos you should download from our uh channel. Please make sure you save the script as a part of the skill um so that all of this is saved. Now, while it's doing that, because what it's actually done here, in order for it to make these API calls, it probably would have written its own uh script, made some API calls. Right now, that it hasn't saved that script inside our working folder here. You can see that it only has the files that we added or the files that it downloaded so that I can actually use this then on repeat again in the future. I'm going to ask it to save it as a skill and in that skill include any scripts. You can see here it says this is a great workflow to capture. So, it's being very very nice. Uh, when you give a channel ID instead of a single video, how should the skill decide which videos to pull? I'm going to say all videos. How much do you want to verify the skill before it's done? So, that's this inbuilt evaluation loop that I showed earlier. I would always recommend running evaluations when it is something that is especially mission critical. Obviously, in this case, since we are just demoing this, I'm going to say just build it because speed is obviously necessary. Where should the oorthth refresh token live? So that the future skill I'm going to say reuse the environment token. And so this is very important. This skill will only work when I'm working in this certain folder. So we have this like two levels here. If I go and call this skill or if I share this skill with someone else, it doesn't give them access to my API. It also doesn't um give them access to my API key because of the way that I've set this up. But you need to be very very careful if you do not set this up the way that I have using this environment file and you for instance just paste the API key in here. I've seen it in many many occasions that then you create a skill and your API key is actually there in plain text saved in the skill. This is absolutely terrible and super super risky and you want to avoid this at all cost. So even after it creates this skill, I'm going to thoroughly review it to make sure the API key or any sensitive tokens are not saved in it. It's purely just referencing that environment file that I created.

All right, now we can see co-work has packaged up that skill nicely obviously based on the answers we gave. And here we have our saved script. So we have the oorthth script. So here it tells it exactly how it should authenticate. And you can see here that there are no uh tokens or API keys because it's basically just referencing the env file. So you can see here appended to env. So it knows to check exactly in the working folder if there is this environment file with all of those saved secrets and these will not be exposed here in the skill. And then we obviously have the Python script to actually go and pull those videos. So let me make sure I save this one as a skill. And now as always what I recommend when you are testing out a skill, don't do it in the chat where you created the skill. You want to make sure is does this skill work on its own? So I'm going to create a new task. I'm going to make sure I'm in the YouTube audit uh folder because that has all the necessary connectors. And now I can just use this YouTube video data um skill that I've created. Download YouTube data for a 9x video or every video on the channel using the YouTube data API. So I'm just going to call this one and I'm just going to tell it to grab our last 20 videos, one dedicated subfolder for each video. I believe this should probably be packaged inside that skill, but just in case it's not, let me drop this here. And then maybe just to show you how even further how you can push this. I'll say, can you also please create a CSV file in the root folder with all of the video metadata and their statistics from these last 20 videos. And one more thing that I want to add because maybe I didn't put it in the skill. We also run like live stream sometimes. We have shorts. I want to make sure that it's only pulling the latest videos which are public, not unlisted. And on our main uh videos feed, please only pull in public uh videos, not unlisted videos or live streams that we have run. And for me this sort of automation is one reason why building out these skills in co-work is drastically different to like creating automations in tools like make or nad because we can create this automation. We create this skill but then it can still use we can use regular human language to adapt the behavior. So we have here this very fixed skill but then we can just provide these loose instructions. We don't now need to go and let's say edit a whole make scenario N8N to change the filters. It's able to do that on the fly. One thing that I would say though is there's 100% room for these tools like make and naden. Co-work is really used to be your co-worker to work alongside you when you still want to stay in control for anything that you really want to happen in the background and maybe at a scale like let's say every time someone submits a form on your website certain action needs to happen and they end up in your CRM. That should always happen in a tool like make or nadn or maybe your tech team are coding that for you. These co-work automations, they obviously only run even the scheduled tasks either when you tell it to run it or your co-work desktop app needs to be open. They are not really made for like instant workflows that need to happen as soon as someone take some action on a website for example.

And we can see just like that. So it's opened up the CSV here. But what I'm more interested in, this thing literally ran in like under a minute. Just a few seconds. This was insanely fast. Let's go and take a look at that YouTube audit folder. So have a look in here. And wow, just look at that. every single video that we have now um by date with all of that data. We have the markdown file with all of the details. We have the thumbnail. We have the transcript file of all of those latest 20 videos. And of course, then we have this CSV file. So, that really shows you maybe like the one learning that I want to take away from this by having co-work actually write this script. Of course, the first time it works, it needs to figure out how to create the script. it might take a little bit longer, but then afterwards, the execution is super super fast. And scripts can also be useful when dealing with connectors. Maybe it's getting some data back from a connector in a certain format and you want it to manipulate that data. Have it write a script for you and then execute the code. And for me now, having all of our YouTube videos saved here in this format with the transcript, with the description, with all of these unlocks a lot of possibilities both in terms of analysis, but also repurposing this content. And in that YouTube API, I can now also get it to actually download the video files and add it to all of these folders. That would take uh a lot longer, but it's literally just an instruction away, but obviously that's like many gigabytes uh to process. So, I'm not going to run that now, but you'll see exactly how and why I've done that in our context system that I'm going to show you in the last build walkthrough of this section.

All right, time for live build number three. And this time, we're going to be focusing on browser automation. And so, as we already covered, Claude does have the ability through its Claude in Chrome extension to actually open up your browser, navigate different pages, click on buttons, and extract data. Now, this is something that you should only really reach to if these first two ways of connecting with tools is not available. So if you've had a look that what you're trying to do isn't available, isn't possible via connectors, it's not possible via having co-work write some code and then run it on its virtual machine, then the last resort is browser automation. So pretty much anything that you can do in a browser now, thanks to co-work controlling your Chrome, it can also do. However, I also want to point out that there is an order to this. So even co-works in this order. Whenever you ask it to do something, it's first going to check, do I have a connector to do this? If not, can I write some code to achieve this? And also, its last resort is using the browser. And that's because browser automation is probably one of the most heavy in terms of token consumption. And it will use up your usage a lot faster than just having co-work. And this is actually a situation I've found myself in recently needing to use some browser automation inside Claude. Basically the situation is as I've mentioned already earlier we run a free live AI workshop each and every week 90 minutes for our community. Also check in the link below if you want to sign up for next week's session. You can see here some of the recent sessions that we've run like co-work for sales building claude skills for beginners co-work for marketing using claude with Microsoft Office and even a deep dive on the new chat GBT agent feature. But basically we recently moved our live uh webinars. We used to run these on Zoom. We now run them inside our new community which is hosted on a platform called Circle. And basically after every session I do need to process the recording because we need to share it for instance with our pro members. We check if there are any highlights that are worth sharing on social media. So we always need to grab the recording and with Zoom this was super straightforward because you could actually do this all using the Zoom API. Now, of course, I've checked with Claude if the Circle Connector, the Circle MCP server is able to do this. So, using the Circle Connector, are you able to list all live stream recordings and download the video files and their transcripts? And basically, then Claude did a quick check using the Circle Connector and said, "Here's what the connector can do. Transcripts? Absolutely. It can list every room and pull the transcript for any recorded session and video files. No, the connector exposes live room metadata and transcripts, but there's no endpoint to retrieve the recording videos URL. So, what can we do? Either we give up and we keep doing this manually ourselves, but there is a better way using browser automation. So, what I've noticed whenever I'm the host of one of these live streams, after the session finishes, we actually get an email from our Circle platform saying the live stream recording is ready. I can download it here directly in email, but there's also a link to the live dashboard. And if we take a look, that link just drops us into the live streams section of our like admin panel of our community. And here we have all of the recordings listed. If I want to actually download a recording, I can just click here. I have to click on the three dots. And this is what we're going to be training Claude to do. And when I click on download recording, what actually happens is this just opens up a new window with a dedicated URL for that video. And I'm wondering is as soon as Claude has this, is it then able to actually download this? So this is what we're going to try and teach it to do in Chrome.

Now this ability for co-work to actually open up your browser and navigate pages. This is not something that is by default built into co-work. You do need to go and install an additional extension. So we have the clawed in Chrome browser extension give you browser automation capabilities. So what this is, this is a Chrome browser extension. So I've got it set up here on my Chrome. Also important to know that right now it only works with uh Chrome and Microsoft Edge. if you're using other browsers like I am for instance is I'm using Arc. So I basically just have Chrome only for Claude to use. I'm not using it in my default browser. Hopefully these will also be added in the future. But once you just head to the Chrome extensions marketplace, look for Claude in Chrome. Then you can install it and you basically have this Claude extension. And this is something that always needs to be open when you want to give Claude a task that involves it opening up your browser. You first need to have Chrome open and you need to make sure you're actually logged in to your Claude in Chrome extension. And I'll also leave a link to the exact Chrome extension in the resources inside our community so you can easily just go and install that yourself.

All right, so let's see if we can teach Co-work to download these videos for us. I'm going to start a new task. I'll create a dedicated project for this. Create a folder. All right, so I've got my download 9x live recordings folder created. And that's going to be the folder we're going to work in. And now I basically need to it's now my job as the context engineer teaching Claude how to actually go about doing this. And maybe one more thing in addition to the Chrome extension also in your connector settings. So as if we go to customize and then connectors also make sure you'll see here clawed in Chrome included and claw in Chrome lets you handle work in the browser and once enabled browser tools are always available to Claude. You need to make sure that this is enabled inside your settings as well as having that browser extension actually installed. And now what I found to be the best way to actually prompt Claude when you want it to do something in the browser is to actually show it. So in this case, what I'm going to do is I'm I've got to my live streams uh page inside our circle community. And I'm just going to click on these three dots so that we see this menu. And what I'm going to do is I'm actually going to take a screenshot. And what I like doing as well is actually annotating this a little bit to make it super super clear for Claude. So I'm just going to point to the fact that it first needs to click on the three dots. So that's going to be number one. And then it needs to download the recording. That's going to be point two. And let me copy in that uh screenshot here so that Claude can see it. Can you please open up the following page in our um circle community settings? And I'll just grab it. give it the URL. Use claudin chrome to then click on the three dots next

To the Microsoft Office uh live stream recording and then click download recording. So here both with the screenshot and in my prompt I'm really telling it the steps that it should do. And let's send this one off.

And straight away Claude has detected that it needs to use its browser automation. So it's going to load up its Chrome tools. Now we see here Chrome browser one. Here we can see Claude is now in control. What it does in Chrome, it creates these different groups. So in this one I had I can actually close the one I had open where I made my screenshot.

Now Claude is in control. Here you can see Claude started debugging the browser. We see the little orange uh mouse that Claude can use. This is pretty pretty cool. That mouse has now just moved over to that section. And we can see now it's already been able to open up uh that menu. And of course there we go. Let me just uh mute this. and how it's actually opened up that video. So, it's telling us done. Clicking download recording, open the recording file in a new tab. The browser is now fetching that MP4, which is the May 21st Microsoft Office stream recording.

Now, what I'm going to do is I'm going to ask it to in the next step see if I can download that to my Chrome downloads folder. Can you download that um video that's now open to the Chrome download folder?

Now this is an important point is when Claude tries to navigate to different pages in Chrome it's going to ask you for permissions and we have a few different options here. So we can allow just for the specific page that it's viewing now or we can say all for this website. So if there is a website maybe like your own website in our case this specific website to our 9x community and we want it to always be able to access this pages I can say all for this website and then I don't need to answer this again and again.

So it says done. I click download in the video player. So Chrome is now saving to your default Chrome downloads folder. Large file. Give it a moment to finish.

Now this is very very important because right now Claude doesn't actually have access to that downloads folder. So what I mentioned earlier when we were talking about the working folder, Claude only has access to the folders we give it. Let's say I want to move over that download into the working folder that I've created. What I can simply do here is I'm going to add a new folder. So, we can add a new working folder in the session. What I've actually set up because Chrome is not the main browser that I'm using, I've created a dedicated folder for my Chrome downloads, and I've actually called it Claude downloads. And I changed the default download folder location in my Chrome settings. So, I know whenever Claude is going to download something in Chrome, it's going to save it here. And I can just attach this working folder now to any session where I need Claude to automate the browser and download files.

I've just given you access to the Chrome downloads folder that I'm using. Can you move that file over to the 9x live recordings working folder and Claude's confirm that's done? The file is now in your download 9X live recordings folder. Let's go and take a look. We see it here already. But if I open up the working folder, download 9x live recordings. And there we have that exact recording.

So now pretty smoothly I've been able to get Claude to download the video files directly using Chrome because we cannot do this via API. We cannot do this via a connector. So the obviously the next step now as you should know by now is creating a skill so that it can do it on repeat because of course doing this step by step with Claude telling it to go and visit this URL, download this file. This is way slower than me just going into my Chrome clicking download recording and then it being done with. But the whole point here is that it's going to be way slower the first time you do it. You tell Claude step by step how to do it, walk it through the process.

Now, I'm going to save this as a skill. And I'm also going to set up a scheduled task so that it actually goes and checks this live streams page. Checks if there's any new recordings that are ready to be downloaded and then downloads them and saves it in the appropriate folder. And this means this is now a task that is off my table. I don't need to go and do this every single week. Claude will just based on the schedule that I give it. So in my case, I know that our 9x lives either run on Wednesday or Thursday evenings, I'm just going to set up a scheduled task to run in the evening at let's say 8:00 p.m. when I know the session will definitely be finished by. It should just go and check this page and then download those files and it will just run in the background.

All right, now for the last one of this main build section. And this one, it's not necessarily a live build, but more of a live walkthrough. And that's because I want to answer one question that I get quite often. Once people start using co-work, they see how powerful it is. But obviously, one big thing is the fact that co-work is really made to run locally on your computer. So, how do you actually work with co-work as a team where you have multiple people all working on files? They all have their own uh local files, but there's obviously files that need to be shared amongst the team that everyone is working on. So, I just want to walk you through the full context setup that we've set up for ourselves at 9x so that everyone has access to the same files, the same folders, the same context, and we're all benefiting from each other's work all at the same time.

So, what we've set up for ourselves is a folder that we call 9x Claude. And this is a folder that sits on my computer, but we are using the Google Drive desktop app. So, this is actually a folder in Google Drive. If you haven't seen it before, Google Drive has a desktop app. And what that desktop app allows you to do is to actually sync folders from your Google Drive to your desktop. So, what we've done is we've set up this 9x clawed Google Drive folder and then it is synced to all of our computers. So when co-work starts working with it, it just sees it as working with local files.

Now within this 9x claude folder, we have a few different things. The clawed MD I'm going to touch on in a second when I actually load this up into claude co-work, but we have two folders. One for context and one for work. So this context folder, we've designed this, this is read only. This is not something that when myself or the rest of the team are doing any work, nothing should be getting saved in here. This is very very deliberate. What is the content that we're making? What are the courses that we are making? What is our sales strategy? All of that lives here in context. But this is very much designed as this is only Claude reading it whenever it needs to figure things out or get additional context. When Claude is actually doing work for us, that all lives here in the 9x work folder. And we have this broken down into different subfolders.

inside the work content folder is where I'm doing most of my work or when I say me is where my Claude is doing most of its work. And here we have this structured for the different content elements that we're doing. So for instance on those weekly live webinars that we're running. I just showed you in the previous uh example how we go about downloading those recordings. That's something that has actually been already running for us for a few weeks. We've got this structure that every single live workshop that we run has its own dedicated folder. And within that folder, we have all of the recording material. So after a live session happens, we get the recording videos, we get the chat transcript, and we also get the full transcript of what happened, which is more than enough that Claude needs to actually repurpose it for LinkedIn, make uh snippet videos from it, um send the summary email to our community about what happened in the live session. It's also where all of the uh visual assets that Claude creates for the sessions are happening as well as the full event description. And for example, on the YouTube side of things, we do the exact same thing for every single YouTube video after it's uploaded. We have the video downloaded here, the full subtitle file or text optimized for large language models to be able to understand it, as well as a markdown file with the overall metadata and the description of that video. And even though my claude is the one generating all of these, this is shared now amongst the whole team thanks to that Google Drive sync.

To make this even easier to visualize here in Google Drive, if I look on my shared drives, there we have that 9X clawed folder. And you can see here the folder structure is exactly the same because it is the same folder. We have the 9x work and 9x context within work. I can go to that content folder and we see all of the exact same things.

So let's take a look in co-work when I actually mount that 9x claude folder and you can see here this is just using that Google drive sync path. So this is the local version of the 9x claude folder and I'll just ask it something like what is the type of work that gets done inside this folder. And what we see straight away on the right hand side is we have these instructions. And now this is actually not empty. This is filled in because we have this claude.md file in all of those folders inside our 9x claude folder. So here we're explaining, hey, this is a Google Drive shared root folder. It contains reference material and working files. Critical rule, never create or save files in this root folder. We want to keep it clean. As you know, when multiple people start working on anything, it can end up getting very, very messy. So what we make sure is nothing goes in this 9x claude folder. If it's workrelated, it goes into the 9x work folder into the appropriate subfolder. And what we mention here is that if there's another mounted working folder, in most cases, that is where the data should be saved. So what someone can do is mount the 9x claude folder and another folder on their desktop. It can they can use 9x claude to basically use as context pull in all the information. But if they're mounting another folder, that is typically where the work should be saved for that session or otherwise it should go into a subfolder of 9x work. And then we're giving it more instructions so it knows exactly where it can find where to save different things. And here we're laying out the folder structure and what exact process it should do whenever it starts a session involving this shared drive.

And if we actually take a look at and what that looks like here just in the finder on my Mac, this is that claude.md. So whenever we add this claude.md file into a working folder, Claude is going to read this first before it starts working. Now we don't only have this claude MD on the root folder. Also when I go down one level into 9x work, we also have another clawed MD there, which gives more specific information about the structure of our working files. And the real beauty of this is it doesn't matter if our team's working in clawed co-work or clawed code. Both of these tools respect this Claude.md. This is actually something that was first created for Claude code, but Claude coowwork respects it the exact same way. And if of course if we take a look at that response, this is the 9x shared Google Drive workspace. Two main areas, context and work based on the skills and connected wired into this space. The actual work that gets done here includes course and lesson production, 9x live events, community and content, marketing ops, internal ops, and you can see this is that whole thing that I hopefully I was building up throughout this course on context engineering. We're now not writing better prompts. This is a very very lazy prompt. I can show you another example in a second. It's about giving co-work access to the right information it needs to get tasks done. for example, for other members of the team that maybe aren't working on content. We will need to blur this next one out, but for instance, here we have a work companies folder with a dedicated subfolder for every single company that we are either working with or potentially working with. Any sales calls that happen, any sales follow-up tasks, any deals, any contracts, any documentation is all saved here inside the relevant companies folder. So Claude always has access it when I say hey what's the latest with company Y it has access to this it also can then connect to Atio our CRM and get even more information that is what context engineering is all about and a perfect example of this here you can see I'm working inside the 9x claude folder and I have a very very simple and lazy prompt what were the highlights from this week's 9x live session on claude co-work for marketing um which will be in the 9x live subfolder it then found that exact folder where it should work. Ran a skill that we have set up for getting video highlights. So, this one actually looks at that long video recording, looks at the transcripts, identifies key moments, and actually chops up the video. And then here we can see co-work just basically delivers these five videos for me to review. And if we take a look at one of these inside the folder, I can just click show on folder. And here we have those five uh clips. We also have like a markdown file or a PDF suggesting potential ways or angles that we could actually um share these clips on LinkedIn. But for instance, if I take a look at this one here, we can see here is that webinar um section, the highlight 54 seconds. And what the skill also does, it al auto automatically adds on our little 9x live branding at the end. And what's very very nice because of this good structure, all of these clips that were generated were saved inside that exact same 9x live folder where we have all of the recording. So for instance, if I check this one, Claude Co-work for marketing. We have all of the assets from my live webinar. I have the description. I have all of the recording files and now we also get the social media clips. All because everything is structured in the right way. We've created the right skills for Claude to know exactly how the work should be done. And this is something that is now accessible for the whole team.

All right. Now, I quickly want to cover a few power more advanced features that maybe didn't get their own dedicated chapter in this course, but are still important to touch on. Starting with plugins. So, here specifically, what's the difference between skills and plugins? Plugins is something that probably more and more of you are hearing about. It's a question we also get asked a lot. When should you use plugins? When should you use skills? What is actually the difference? And the most important thing to know here, these are not two completely separate concepts. they actually belong together. So obviously by now you should know in detail what a skill is. Also how to create it. A skill simply put is how Claude should do a specific task. The instructions of how Claude should execute on a specific task. How Claude should work with a plugin. This is a sharable setup for Claude. And a plug-in can actually include or in most cases will always include multiple skills. So let's say you start working, you're working in marketing, you build a skill for writing newsletters, you build a skill for maybe creating newsletter images, for creating social posts, for creating LinkedIn, and you want to share this now amongst your team. Instead of sharing each one of those skills individually, you can create a marketing plugin, combine all of those skills together into one plug-in, and then make it sharable. So skills are the actual definition of how something should be done. A plug-in is just a way of combining several skills and sharing them. However, plugins also include some additional information. So, the plug-in will also immediately include what connectors, what tools does Claude need to access in order to execute any of these skills. So, let's say in your marketing setup, you create your marketing skill. You would include in the plug-in the connector to your CRM, the connector to maybe your ads management platform, your content management platform, whatever that may be. here. It means when someone installs the plugin, they're not just getting those skills, they'll also have those connectors added to their account. Then they'll obviously just need to go and log in and connect to those individual connectors, but they're already nicely added inside the claw dashboard. There's also some more advanced features that plugins can contain like agents and hooks, but again, that's probably a bit too advanced for uh this current course. So the way that you can think about it is skills teach Claude how work should be done and plugins package skills together with tools and optional automations so the setup can be easily shared and installed.

Let's go and take a quick look at one or two plugins. So here in the customize section we have um of course skills and connectors and then we obviously have some plugins and we've already started creating a bunch of plugins for ourselves here at 9x. Another benefit of plugins is the skills that are a part of plugins will work across not only claude co-work but also claude code. So if you're someone that's using both platforms, plugins is maybe something that you should um reach to. But the one I want to point out as a demo is for instance the Figma plugin. So if we click here on browse plugins, we can see this plugin marketplace and we have this tab anthropic and partners. Some of the different software that we're using each and every day like Figma or Air Table are starting to create their own plugins to help their users get quick access to sort of best practices within their tool. We can also see like Cloudinary, Zappia, Bright Data, a lot of different um plugins here. If we take a look under the hood of the Figma plugin, see we not only have skills, we also have connectors. So, anyone that adds this Figma plugin immediately gets the Figma connector. they don't need to separately go and click plus and search for it here. By installing the plugin, the connector will automatically be added to their account. So, if you're someone that is maybe you're an agency owner setting up Claude setups for companies or if you're a Claude admin, plugins are very very useful. For instance, a new employee starts on day one, they install the necessary plugins, they already have all the right skills and connectors set up and ready to go.

Next up, live artifacts. And this is a feature that I'm starting to use more and more inside co-work whenever we ask it to work on some data analysis type tasks. We can ask it to create a live artifact. And that's literally all you need to say if you're giving it access to some data, whether that's with local files inside folders or your connectors. So for instance, you could connect your Stripe account through the Stripe connector and have it build an accounting dashboard for you. So what you can do is just literally ask it to do some data analysis and then tell it to create the live artifact. Now, here's an example of a live artifact that I've set up for some data coming from a Google Ads account. And this dashboard was set up just using a bunch of CSV files that I downloaded from Google Ads. So, the campaign performance report, the keyword performance report, and so on. And all of those files can then be used to create a dashboard like this.

Now, one very big bit of news. Up until recently, these live artifacts that you set up in Co-Work were only available to you. You couldn't share them. But a couple of days ago, I noticed this little share button. So now, if you're on a team or enterprise plan, you can click share and create a link that anyone in your company can view. Now, they need to also have access to co-worker in your co-work account. That's how basically Anthropic is limiting who has access to them, but it seems like then a very very secure way of sharing data amongst your company. If you are on an individual plan, I believe you can choose to actually when you click share, publish these artifacts to the web and then share that link out publicly. But for team or enterprise, it's only available to people within your organization, which is super super nice. And the way that you edit these artifacts is simply the same way you create them. If I click here on chat, it'll take me back to the very first conversation where I created this artifact. So now if I want to make any changes I just give it the feedback here.

And the final power more advanced feature that I want to touch on. We did already touch on this right at the start of this course but just briefly is parallel agents or sub aents. So this was the example that I walked through earlier on in the course. If we're asking co-work to for instance research these five competitors instead of doing this one at a time one after the other like the regular chat claude chat would or any other chatbot it can spin up sub agents parallel agents. So for instance it's going to create one sub agent for competitor number one and give it all of these instructions on what type of analysis should be done. Now I mentioned I would talk on this in a little bit more detail. So, what's really going on under the hood is when you are working with co-work, you are working with one agent. One agent runs your co-work session or task. It plans out what task needs to be done, fills in the process list on the top right hand side and starts working on the tasks. And all of that happens inside the one context window. Now, when that agent decides that it should spin up sub aents, so if you give this prompt, it will probably be smart enough to spin up sub aents. So what happens then is every single sub aent works in its own context window. That main agent does not see any of the research each sub agent is doing. And so what that main agent sees, it doesn't see all the steps that the agent took. It just sees the agents final summary. So in this case, what I would expect is all of these sub agents are giving back to the main agent their report findings. The main agent is then comparing all of them and creating the final analysis. Now, as I said, there is built-in functionality in co-work so that it automatically creates these sub agents or decides when to do it. But you can also influence that in chat. For instance, here if you really wanted to make sure that it use sub agents, I would just add an additional line to the prompt saying use one sub agent for each competitor or run these competitors in parallel. In parallel is sort of the keyword to trigger off these sub aents.

Now, sub agents can also be used inside skills which is very very powerful. I'll give you an example here of that one that I ran earlier for getting the video highlights from our 9x live recordings. What we have is we have the skill markdown file how the process it should go for taking a transcript, taking the video file and creating those um highlight videos for our social media. We then obviously have in the assets that outro video so Claude knows exactly what video to put at the end of the snippets that it creates. But then we have this agents and we have two different agents that we have created here. And what this means is when this agent runs as a part of the task, all of this happens outside of the main context window. And we have one agent which is a visual analyzer agent. So what it does is because I want co-work to pick the best potential snippets for social media, it shouldn't be based just on the transcript because the transcript of a video is obviously missing a key part, the visual. So, what we've told it here is that when it's identified some potential clips that could be highlights based off the transcript, we then go further and hand off some work to this visual analyzer agent. What that one will do is actually go and use a tool to take screenshots of every single second of that maybe 30 40 secondond video that we think could be a useful highlight. So, it not only has the transcript, but it also sees what's happening visually on the screen. And then it reports this back to the main agent who can decide which highlights are worth sharing or not. Now again here I didn't set any of this up manually. These agents are simply set up using that skill creator skill. And in many cases when you create a skill, Claude will just create these agents for you. If there's ever a moment where you think, hey, this would be better off using a sub agent. Just use the skill creator and ask it, do you think sub agents should be used in this skill? And it will give you an honest answer. And if yes, it will go and set these up for you.

All right. Now, let's talk usage and limits. And this is something very important to understand in co-work. It is also a bit of an area of contention. Probably if I do hear one main frustration around using claude, particularly claude code and claude co-working limits and are basically then not able to work for a certain time. So, let's unpack this. First of all, I want to head into co-work and I want to show you where you can actually keep an eye on the different limits and how much of your usage you've actually eaten up. If you go down to the bottom left to your profile and then click on settings. And now here in the third tab, we have usage. So, let's click on that one. And now, the first thing I want to unpack is that you have different levels of limits when it comes to your usage. So, the one that you will probably hit the fastest will always be your session limit. This is the same across all of Claude. Whether you're using Claude chat, Claude co-work, or Claude code, they all count towards the same usage limits. And you have this session limit, which is basically a 5hour rolling window. So within a 5hour period, there is a certain limit of usage that you're allowed to use. And now you can see that um my window, my session limit quite recently um reset, more just a little over an hour ago. And therefore I only have 8% used. And as I start using this, this is something that you will see start filling up.

Now, additionally to your session limits, you also have weekly limits. So this is then how much you can use in a given week. Now when it comes to the actual usage that you are allowed, you can see here they're only giving it to you in percentages and your exact usage limit will depend on what plan you are on. If we take a look at the um pricing model of Claude, if you're an individual user, you have these three options. Free, pro, and max. As you already know that if you're on the free plan, you do not get access to Claude co-work or Claude code. So my assumption if you've just sat through this multiple hour co-work course, you're probably at least on the pro plan. And you see here that the pro plan obviously you're paying for it has more usage than free. And then you have the option of the max plan. And in the max plan, you can either get five times or 20 times more usage. And this is the exact same when we look at team and enterprise. So on the team side of things, you have a standard team seat which is basically equivalent to pro with a bit more usage. And then you have a premium seat which is five times more usage than the standard seat. So all they're telling you is that the more expensive seats, you get five times as much usage. But while we see that then the more expensive plans have more usage compared to this one. What we're never given is an exact amount of usage that we can currently use in terms of the number of tokens. That is not something that Anthropic are sharing. And mainly because that this usage that you get, it is something that fluctuates. And this became very clear a little over a month ago when Anthropic became super super popular. a lot of people and a lot especially a lot of companies started jumping on Claude specifically Claude Co-work then you basically have a lot more users on the platform anthropic only have a certain amount of compute that they can give out and therefore it meant that each individual basically received less usage but this is not something that they were really communicating that much about but then at the beginning of May so earlier this month and this is how you can tell that this usage is something that fluctuates basically signed a compute deal with SpaceX, meaning they could provide more usage to their users. And what did they immediately do? They doubled the amount of usage that every single individual could use on their clawed platform. And this was for the paid users. So you can see that usage is something that fluctuates. But what's always consistent is that the premium seats, so the max seats or the premium seats in a team plan have far more usage than the standard or pro seats. So me personally, I am on a premium uh team seat and I very rarely hit my limits and really if you compare how much more productive this software is going to make you, I really think it is a worthy investment.

Now what happens when you hit these limits? So for instance, if you're going to hit the session limit, let's say if I hit this session limit with more than an hour to go, basically then I need to wait. I cannot use co-work anymore until this session limit resets. However, you do have one additional option. Basically, you can upload an amount of credit to your Claude account and then turn on usage credits. What this means is that if you ever hit a session or a weekly limit, you can continue using Claude and then it's going to eat away at any credits that you have uploaded. And this is priced the same way as if you're using Claude via the API. If you're a company admin, you can see which of your team members have actually eaten up any of this credit. And you can also set monthly limits per team member.

Now the main question I get asked a lot is okay how can I avoid hitting my limits and what are the best practices around this. I'm going to touch on that a little bit in the next chapter as well where I give my general best practices and recommendations. But for now what I can show you is I I'll again use that um context skill that I showed during the context engineering part of this course. And this is a very useful tool that we can use to help us understand how we're consuming our limits and how to avoid hitting them too quickly. So one important thing to remember when you are working in a task in co-work you start filling up this context window and we see here in this current conversation that I had running I'm at 100k tokens out of a possible 1 million. So previously for context most of the anthropic models had around 250k tokens that we could use until it would fill the context window and then what would happen is that this context would compact. What Anthropic would basically do or what Claude would do is make a summary of the conversation, compact the amount of tokens in the window and continue the conversation. Now, they've increased that amount to 1 million, which is super super useful. It helps us provide more context in the window, but it also means we're going to start eating up our usage quicker. And that's because one thing that a lot of people don't realize, right now, I'm at 100K tokens. For every message that I send, all of these 100k tokens are basically processed again. So when you have very long running conversations, let's say I have a conversation where the token count is at over half a million. If I keep continuing having back and forth messages in that conversation, I'm going to eat my usage a lot faster. So that is why one of my main recommendations here when it comes to usage is try and avoid these longunning conversations. If the next task that you want co-work to work on doesn't depend on something that already happened in that conversation, just create a new task. This is one mistake that a lot of people make and I'll touch on that a little bit uh later on.

Another thing that I cannot stress enough for helping you save on usage is creating and leveraging skills. What will happen in a lot of times when you are creating skills is the skill will write a script for you. You can see it here for that use case when we pulled in our YouTube data or when we're creating the live assets for our webinars. Basically, when you create a skill, co-work does the work once in the process of you setting up the skill. Yes, you will use quite a lot of usage because co-work needs to figure out how the task can be done. That is the beauty of these agents. You give it a task and it will basically figure out the different ways, do some trial and error. In that trial and error, it's consuming quite a lot of usage. The thing is then once it has actually figured out how something should be done, save it as a skill. All of that trial and error will then be put here into the skill markdown file into any scripts that it create. And the next time you run that process, it becomes a lot more efficient. If you do not save these skills and then you keep making co-work do that same discovery over and over again, remember that every single task that you give it has no context about anything that happened in previous tasks. it will need to do that discovery again and will end up burning your usage faster.

And the last one on usage, important to remember that co-work is an agent. It's spinning up a virtual machine every time it runs. Co-work will use more usage than regular chat. So, as powerful as co-work is, and as amazing it is, if you just have a question that you want to get an answer to, do not use co-work for that. Use chat. If you just want to ask a question and get an immediate answer, keep leveraging chat for that. And I do it all the time when I'm doing a bit of research. I'll ask some questions in chat. I get my responses a lot faster and it's a lot more efficient on usage. Now, when it comes to based on that research, if I want to then take that over and work on a task in co-work, what I'll often do is I'll just tell the chat, hey, I now actually want to implement this in co-work, can you draft the prompt that I can use? And that is a really, really good workflow leveraging chat for discovery and then implementation in co-work.

All right, and we have made it to the last chapter. Common mistakes and best practices in co-work. My final recommendations and tips. First of all, I just want to say thank you so much for sticking through this course. I hope it's provided you with a ton of value. If it has, it would mean a great deal if you can give this video a thumbs up and also subscribe to our channel. Plenty more co-work content coming your way. So, let's get into it. The common mistakes that I see all the time for people using co-work and also some of the best practices. So, here we go.

Best practice number one, use templates and references. So, if you already know how the work should be done, and this is the case in a lot of things that you're going to hand over to co-work probably right now you're writing some proposals yourself. Maybe you're creating some emails yourself. Maybe you're creating some assets yourself. All of those things that you're already doing yourself are perfect candidates for co-work. But don't make co-work do the work from scratch. Provide it with the templates. If you want co-work to create invoices for you, provide it with the template of what the invoice should look like. And if you want co-work to fill a sales proposal document based on a call transcript and you can use co-work in doing this as well. Draft the ideal sales proposal template in Google Docs in Microsoft Word. Then provide it with that template that it should update. You really need to show Claude what good looks like.

Next up, when you are working with connectors, so when you have co-work connecting to different tools, always provide the tool IDs or links. And this is another way that will save you a bunch of uh usage and credits. So let's say I want co-work to work in a certain Google sheet. I could of course just say, "Hey, could you find the monthly reporting from last month spreadsheet?" It is going to use up a bunch of tokens in searching for that name because you've just given it a similar name to what it might have. It needs to actually go and find it. much better way if you can just go and grab the URL of the spreadsheet, paste it in and say, I want you to work in this exact spreadsheet. This is the same then if you want it to post in a Slack channel, give it the Slack channel ID. If you want it to work in a specific notion database, go and grab the notion database. And in most cases, the easiest way to do this is just grabbing the URL because in a lot of tools, that specific ID is already in the URL, and that link is all Claude needs to understand exactly where you want it to work. Apart from being more efficient, it also makes sure that Claude is not doing something that you do not want it to do. We know how lazy we can get with file naming and and this sort of thing. And it's very easy for Claude to make a mistake and start updating the wrong file or working in the wrong folder. We want to avoid this at all costs. So when you are using connectors, always provide the ids and links of where co-work should be getting its information from and also where its work should be done.

My next best practice and another common mistake I see is people see that there is a connector available for a certain tool. So they just jump in and use the connector straight away. What I highly recommend and this is really one for any sort of reporting or analysis use cases is do the work first on local files especially if a connector can't provide the context or take the action. Some people see ah this connector can't do what I want it to do therefore this is not a viable use case in co-work and that's absolutely incorrect. Take an example of Google ads for instance. Right now there isn't a Google ads connector. So I'm not able to pull in any Google ads data into co-work. But that doesn't mean co-work can't help me in managing my Google ads. What I would simply do is go into Google ads. You can even schedule these reports. Download the five or six main reports manually first time with a CSV. Drop them into a working folder and then have co-work on those files. And this is even a best practice if the connector exists. Because when co-works on the local files, a few things happen. A, it's able to do its analysis a lot faster because it doesn't need to pull in that data live. It sees the CSV, it sees all the column names, and it can get to work on writing its scripts. Another great benefit of doing the work locally first is Co-work sees all the data that it has access to. In some some cases, what will happen is the data that we are able to get via an API or via the connector isn't the same as what we're able to get in our manual exports. It's frustrating. It's annoying. I wish it wasn't like this, but it is the case in a lot of tools. So, when you get co-work to first work on local files, CSV exports, Excel files, what's great, it knows, hey, this is all the data I have available. When it then goes, if you actually get it to run an analysis and say, hey, I now want you to connect to the connector and pull in the data live and see if you can run this on repeat, what it's going to tell you is it's going to say, hey, these are all the necessary fields that I do not get in that connector. as opposed to if you just get it to connect to the connector straight away, it's not going to know that those fields even exist. And so, it's not even going to flag this to you, and it's going to tell you, hey, this is a full analysis, even though it's missing some key information. Another big thing to watch out for on the connector side of things is some connectors only pull in a subset of the data. So, let's say you have a 100,000 contacts in your HubSpot. If you go into a HubSpot and just download the data, maybe make it miscellaneous, so you don't include like email and first name and last name, but you want co-work to analyze the attributes. If you do that in a CSV file, it's going to be super fast. It's going to know exactly everything about the 100,000 contacts. Now, if you do that with a connector, maybe co-work only pulls in the first thousand and doesn't realize that there's more. You specifically need to tell it, and it's going to say, "Hey, this is the findings that I have about your contacts," but it's only looking at a subset of the data.

Next up, I think you should know by now, this is one of the most important. Create skills as you go. Whenever you work with co-work and it figures something out, it's doing some debugging in a tool. It figures out, hey, this is how I can get data from notion or this is how I should analyze these certain files. Make sure you're saving this as a skill so that you doesn't need to do that discovery again and again. We sort of just covered this in the usage and limits chapter, but this is one of the best practices that you can get into. And another common mistake that I really see people do is they get excited about creating their skills and they try and create skills straight away. I always recommend run the process with Claude first, then save the skill. Don't just open up the skill creator and say create a skill that does X, Y, and Z. First run the process together with Claude. Once you're happy with the results, then create the skills as you go. And also this as you go also means you should be iterating and updating on your skills.

Next up, and this is the one that really will save you on your usage. We again we just sort of covered this, but avoid having long conversations. Whenever the topic shifts, start a fresh task. And I'll give you a specific example of this. It's something that we also covered a little bit in the course is I use co-work all the time in helping me set up our weekly 9x live webinars and workshops. Whenever I'm working on a new workshop, I create a new task. This is one of the biggest u mistakes that people make specifically. This was coming a lot from the early chatgbt days and I was really guilty of this myself. I'd love to know if you're also in the same boat that you'd basically have this super longunning conversation with chat GBT because you figured something out once. So let's say I had a conversation with Chat GBT about preparing webinars. Then every time I would work on the next webinar, I would just go back to the conversation because back then we didn't have skills, we didn't have custom GBTs or we didn't have projects. And this created a lot of bad habits for a lot of people. And now I'm seeing a lot of people bring that bad habit over to co-work. They get co-work to work on their webinar for this week. They get some good results and then when next week's webinar comes along, they use that same conversation. What you want to be doing is creating the necessary skills and then start new tasks because even though both are involving a live webinar, these are different topics. you creating the next week's webinar on a different topic has nothing to do with what you did last week. And so, wherever you can apply this, always start a new task, even if it's about the same thing, make sure you're leveraging skills and avoid these long conversations.

And now, my final recommendation when working with co-work is delegate work and check back. Now, a lot of people are used to coming again from the chat days, whether that's chatubitty or claude chat. Both of these give you an answer quite quickly. So, the benefit of a chatbot is you're getting these fast responses back and forth. With co-work, as you may have seen already throughout this course, it is a lot slower. That's because it's actually doing the work. It takes its time. It thinks through the process. What you need to do is get in the mindset of like imagine you are delegating a task to a colleague on Slack. If you message someone on Slack saying, "Hey, can you take care of this for me?" You're not sitting there waiting for them saying, "Hey, are you finished? Are you finished? Are you finished?" You delegate the task to them and then you go and do something else and then maybe an hour later you think, "Oh, maybe I'll check back. I haven't heard from Sam if he's finished with that task." And you need to treat co-work the exact same way. You should be delegating task to co-work, not just sitting there and watching at work because it does take a little bit of time. It is a lot slower than regular chat. So, what I try and do in the mornings or like break these up into hour blocks, what are the next three or four things that I can delegate to co-work? I'll just go through delegate thing one then move on to the next thing delegate number two delegate number three and delegate number four and then you can go back and check on the progress but do not treat it like a chat and do not sit there watching at work it is a trap that you can fall into but this is really a conversation that I think a lot of people are having this really does mean a different way of working is probably a lot more multitasking than we're used to and this is something that different people do struggle with so really try and pay attention to this one that you're not just ending up sitting there watching co-work uh spin the wheel for you, try and delegate and then move on to something else.

All right, we got there. That is everything you need to start using co-work at a serious level. But now, if you want some inspiration on where to apply it and see how I'm using it every single day, click this video right here where I walk through eight use cases I'm running every single week. From scraping competitor ad funnels to creating formatted reports for our clients, these are the actual skills I rely on. And every single one of them is free to download.