Transcription
A key trend in 2026 is the move from chatting with AI to letting it execute the work for you. One of the front runners in this space is Claude Co-work, which is essentially a product that lets non-technical people tap into agents that automate the work away for them. It can do things like creating reports and presentations, organize files and emails, research and analysis, or general work automation.
And one thing worth calling out which I particularly appreciate is that this product comes from Anthropic, who is known to put AI safety at the center of everything they build. So, compared to a lot of the open-source alternatives, the risk of something going sideways here is meaningfully lower.
And by the end of this video, you will know exactly how to get the most out of Claude Co-work, even if you're a complete beginner. And because I respect your time, I will highlight the key topics in this video so you can skip ahead if you're here for something specific.
If we haven't met before, I'm Ali Salam, and I currently work as a director in a tech company. And on this channel, I'll help you turn tech and finance into your personal advantage.
In order to access Claude Co-work, all you need to do is create your account on claude.ai. Once you log in, this will be your landing page. And you just need to navigate to the download symbol at the bottom. Once you click it, you will have a download section here for Co-work. And once you've set it up, it's going to look something like this. It looks fairly similar to the view that you had in the browser, but this is the actual cloud application. And it has way more capabilities.
The important thing to be aware of is that you have three tabs at the top. You can see Chat, Co-working, Code. So, Chat is where you would chat with Claude, but it won't execute any work for you. So, basically, you can expect this to work similar to ChatGPT. If we shift to Claude Code, this is designed for developers. It writes code, builds software, runs scripts, etc. And then we have Co-work in the middle, and that essentially bridges the two worlds.
What's really nice is that the UI is super simple, right? It looks the same as any other chat window. And honestly, the time investment to learn how to use Claude Co-work is very small, but the capability or the skill cap of what you can do if you master it is very high. So, it's really going to be your ambition that sets the limit of how much you get out of Claude Co-work, but even with very little input, you're going to get tons out of it.
So, let's just cover some of the basic principles that you need to know of. You give Claude access to certain folders via this button, and then you execute tasks on your behalf in the chat, as you would normally do in any of the other chatbots. You have the model selection on the right-hand side here. So, Sonnet is basically the one that you can use for generalist tasks. Opus is what you would use for more advanced tasks, and Haiku is what you would use for quick answers. Honestly, when I'm using Co-work, I basically just stick to Sonnet 99% of the time. If I'm doing something super advanced, I might switch over to Opus, but the rest of the time, I'm basically just plugged into Sonnet, and that works great.
For the extended thinking at the bottom, I don't ever toggle it off. It consumes a bit more tokens, but then I am also sure that I don't need to toggle it on and off depending on the complexity of what I'm asking. It just saves me more time. I would just keep it on unless you are very picky about the token cap.
Under the plus sign, there's one thing you need to know how to use, and that is the connectors. So, we will be using this quite a lot in the video. I will return to this. But what you need to know about the connectors for now is that it allows you to plug into external tools, as you can see here, and use workflows outside of Claude Co-work. But again, we'll return to that later in the video. For now, that's all you need to get started. So, let's get the show on the road.
So, as I mentioned earlier, what makes Claude Co-work different than the other AIs is that it not only chats with you, but actually executes work on your behalf. Let's put that statement into action by letting Claude Co-work go out and find us a set of reports on a topic of our choice. Let's do AI adoption, and then build us a report based on the sources that it finds.
So, we'll start off by selecting a folder. Once you've selected your folder path, all you need to do is confirm the folder selection. And then it's going to ask you whether it's allowed to change the files here. So, you can either click "Allow once" or you can click "Always allow." I will do the lazy option here. Then we'll just drop in our prompt. So, let's say, "Search online for the most recent reports from MBB, big four on the enterprise AI adoption, then create a professional report, source selection at the end." All right, let's fire this one away.
Once you actually put it to work, you will notice a couple of things happening. So, on the right-hand side, you will see that it outlined the plan. Then you see the folder path that we've selected. And lastly, you're going to see the context window that it's using. So, it's going to use the connectors, which is in this case the web search. We will see more advanced MCPs later in the video. Then we have the skills that it's putting to work here. So, because we asked it to do a report, it's going to pick up the docs skill. I don't even know how to pronounce this. Docs. And one thing that is good to know is that you can have multiple agents working at the same time on different tasks. All you need to do is open a separate window, fire off another task, and you will see them stacking up at the bottom here and working concurrently.
All right, so the report seems to be finished. Let's go to our folder, and we can indeed see that it's built up a Word file here. Let's pop it open and look at that. I think it looks pretty good. It's starting with an executive summary, and then it's breaking down the report into its various chapters. So, we don't need to go into detail of it, but I think this looks like it's drawn a couple of insights. It structured it up into digestible points, and at the end, it has indeed left a source selection. I think I misspelled it. I was supposed to say "source section," but hey, potato, patata, right?
Anyways, let's go back into Claude and get it to build out a presentation for us based on this report. So, let's just simply say, "Build a corporate presentation based on the report." And here we can see that it's now pulled in the PowerPoint skill to augment the material into a deck. All right, so it has finished generating the presentation. We can see that it built out 13 slides, and the structure of each slide is outlined here. Let's go back to our folder and see. And we do indeed have a PowerPoint deck here.
Now, let's plug that open. Okay, so it has clearly created a deck. But let's be honest, it doesn't look great, right? Like, some of these slides look okay, I guess, but it's not a great deck. And there's actually a specific reason for that. And that is because we're now kind of pushing the boundaries of what Claude is capable of doing completely alone. And if we want to enhance this, we would need Claude to connect to external sources or tools that are specialized for the task at hand, in this case, building presentations. So, that's what we're going to do next.
Generalist AIs will cover the majority of what you want to do, but every now and then, you will bump into tasks that a generalist AI either cannot do, or as we just saw, it can do the task, but the output is not at the right quality. And for both of these two scenarios, we need to turn to external tools to solve them properly. And the connections to external tools exist in Claude by default. And you can access them through the plus sign. Just click, and then you move to the connectors tab. And here you'll see all the connectors that I currently have active. And by the way, connectors and MCPs (Model Context Protocols) are used interchangeably in this context. It basically just means that you're connecting to external tools.
So, at the most basic level, you can see that I have Gmail active here. When it's turned on, Co-work can read, organize, and send emails on my behalf. But if I turn it off and then ask Co-work to do the same activity, it will tell you that it's unable to do so because the connector is not active. And it's going to tell you to activate it before it can proceed.
Now, there are actually a lot of connectors for you to use. So, if you go down to "Manage Connectors," it's going to bring you into this view. If you go to the top and click the plus sign and click "Browse Connectors," you will end up in the library. So, you can see that there are a lot of connectors in here. And the way I would advise you to use this, not to get lost in all the stuff you can do, is use the filters at the top. So, you have the categories up here that you can use to filter down the selection, or you can just use the search button because this is going to essentially search across their descriptions. So, if you want to, I don't know, connect to Jira, then you can just write "Jira," and it's going to filter down to the Atlassian MCP.
And once you found the connector that you want to use, all you need to do to activate it is just click the plus sign, and it will usually take you through an authentication process for whatever you're using. So, if you, for example, connect your Gmail, then it will ask you to log into your Gmail and agree for Claude to connect. And once it's set up, you will find all your MCPs and connectors here, where you will have the ability to toggle further permissions. So, if you, for example, are on ClickUp, which I use to project manage the workflow when I'm creating my YouTube videos, you can decide what Claude is able to do without asking you. So, then you can just click this "Always allow" button, or you can move it to "Needs approval" or "Requests permission." And if it's something that you're particularly sensitive about, you can just block that activity from ever happening. So, when you're connecting to external tools, this is the view that you can use to always stay in control of what's going on outside of your Claude environment.
Now, circling back to our presentation, we weren't very happy with how that one came out, right? So, let's just use one of the connectors called Gamma. It is a specialist AI for AI-generated presentations. I've made a video on it, maybe some of you have seen it. And all we need to do is basically toggle it on. And then we can prompt Claude to use Gamma to create the presentation again. So, let's just say, "Use the Gamma connector to build a presentation on the enterprise AI report just generated." And now we can see that in the context window, it has indeed plugged in the Gamma MCP and it's confirmed that here the Gamma connector is already available. Let me feed it the full report content.
Okay, so it seems like the report is ready. Let's pop it open. Okay, so before we actually go through the deck, let's just look at the old one for reference. So, this is what Claude created, and you can see that although the text is there, it seems to fail on a lot of points. So, for example, it's just plugged in a line here in the middle of the title. Doesn't seem to realize that that's not visually appealing at all. The fonts are a bit weird. Doesn't seem to be a clear structure around how it's telling the story, etc.
If we now move over to Gamma, it looks like this. Even at first glance, the first page looks much better, right? Let's move down. Executive summary. I think it looks a bit light, but still cleaner. Yep. Yep. Yep. So, you can see this looks much more professional. This could actually be a really good deck given some more work. And when you use Gamma, you can actually steer the model in how you want it to build out the deck a little bit more than I did. So, if you want it to be a little bit more prescriptive or follow a certain structure, maybe outlining the straw man with the title headers and somewhat of a rough structure on the content that you want on each slide, it will follow that fairly well. But this looks much better than the Claude-generated presentation.
Claude skills deserve more time than I will give it in this section of the video. And I have made a full video on how to use it. I'll link it up here for anyone that wants to delve deep, but I'll cover the summary here. Claude skills, for anyone new to the concept, are instructions that Claude loads dynamically when executing tasks. This could be things like analyzing data according to your specific playbooks, applying brand guidelines, or general work automation. You'll find the skills section by clicking "Customize" and then clicking "Skills."
Now, there are two types of skills that you need to be aware of. First, you have the Anthropic skills, which are the out-of-the-box skills that come with Claude by default. Then you have custom skills, and those can either be skills that you generate, or you can load skills that someone else has generated. Often times, those are referred to as community skills. And we're going to create a skill of our own today just so you can see how that would look. And the way we create a skill is actually not much harder than just asking Claude to create it after we describe what we're trying to achieve.
So, I will drop in my prompt that I will use to create my skill today. It looks like a lot of details. It's just because I have created quite a few skills, so I know what I'm looking for, but it does not need to be this detailed. It could just be a couple of sentences of what you want to achieve, and then Claude will help you refine the rest. But let's just quickly go through what I have in here. So, I will essentially ask Claude to create a skill that applies my YouTube brand theme on either presentation decks or reports. So, it's going to start off by explaining what we're trying to do: "Create a Claude skill called Ali's YT Brand Skill that applies Ali's YouTube channel brand styling to documents and presentations."
So, first, you have a brand identity. So, you have the color palette of beige and Brunswick green, and then you have the topography. So, it's going to try to use Arial if that exists, and if not, the fallback is first going to be Helvetica, and then it's going to be sans-serif. The next section is scope, and this essentially outlines how it is applying the theme into the Word document and PowerPoint presentation respectively. You can go all the way down to details here of explaining the font size, and you know, you could dump in the details that you want in here. This is just the level that I placed it at for this round.
After that, you have a section for skill behavior. I have broken down this section into two kind of subcategories. The first explaining when it should trigger. So, when the user asks to apply "Ali's brand" or "Ali's YT brand," etc. But then a secondary subcategory explaining when I don't want it to trigger. And the last two sections are basically implementation notes and a skill description.
So, with that said, let's do the same thing as we talked about in the beginning. Let's set a working folder. And a good practice is to separate every project into a separate folder. So, I've created "Ali's YouTube Brand Skill." I'm selecting that folder. "Always allow." And that's basically it. And the reason why it is this simple is because there is a skill to create skills. So, you can see that it actually pulled that skill in right here. It's called "Skill Creator." It's applying that skill anytime you ask it to create a skill. That's why it doesn't need to be much harder than just explaining what it is you're trying to achieve. And in a second, you will see that it will also trigger an evaluation loop, which is super cool.
Okay, so we've received the evaluations, and the way you should read this is it's essentially comparing, like, when I apply the skill, how well do I perform versus when I don't apply the skill for the same task. So, we can see that it's created a report in the first evaluation, and it looks like it scored six out of six in both scenarios, which I find is a bit strange. In the second scenario, it created a PowerPoint, and in this version, the skill got six out of six. The baseline model got five out of six, and it's explaining why here: "Skill wins, baseline missed the beige slide background." And then the last evaluation says, "Prompt engineering docs," who created another report, and here we have the same outcome. So, the skill wins, baseline used white background instead of beige.
So, let's actually take a look at some further details in the evaluation. So, you can see the prompt that I created in order to run the evaluation, and you can see the output file. We can take a look at that in a second. And here we have the formal grades. So, if we download this, can pop open the file itself. And I'm not very happy with this, right? It's not using a beige background. It is using the green fonts, but it's definitely not using the beige background. Now, honestly, I'm not sure if there is an augmentation happening here because I use dark mode in my Word file, but I'm going to tell Claude that I'm not very happy with this. So, let's give it the feedback: "I don't see a beige background in the output file. Please ensure that is fixed in the next iteration."
Let's actually take a look at the PowerPoint deck as well. Okay, so this looks much better, right? So, this actually followed the instructions that we gave it. It has a beige background and it has the green or Brunswick green, I should say, color on the fonts. Let's add the comment here: "The PPTX meets the expectation." The way the evaluations work is it's going to give you the task with the skill and the task without the skill. So, this is the first evaluation with the skill. This is the first evaluation without the skill. This is the second evaluation with the skill. This is the second without the skill, etc. So, what we just did was basically plug in the feedback on the evaluations where we had a skill applied. So, with that said, let's plug that in.
Now, let's go back and tell Claude, "I am finished reviewing. Please apply the fixes." All right, so Claude has now fixed our requested adjustments. All we need to do at this point is click this "Copy to your skills" button, and it will automatically transfer it over to our skills repository. So, let's fire that off. And it's telling us that it's copied it over to the skills. Let's go there and have a look. And here we go. So, once it's here, you can actually see the details of it. This is how the skill itself looks. This is the content of the skill file, which is just a markdown file.
Now, let's put this one into use just to see how it actually works. We can go back to our AI adoption chat. And then we basically just say, "Apply the Ali's YT brand skill on the enterprise AI report you created, and save it as a new version." Let's fire this off. And we can now see that it has indeed pulled in our newly defined custom skill. So, let's see how this actually turns out.
Okay, so it's finally done. Now, for reference, this is how the original looked. Or actually, let's scroll up a little bit because I have dark mode on, so I don't think it does it justice. Let's pop it open here instead. And you'll see that with a white background, this looks a little bit more corporate, right? Looks pretty good. Now, let's scroll down and see how it applied our brand theme onto the same document. And as we can see, we have a beige background with green text. So, this does indeed follow our brand guidelines. So, anytime I will now create a PowerPoint deck or a report that I want to comply to my brand themes, I can now apply this skill to have a coherent and consistent structure of the material that I create for the channel. And you can of course apply the same in your corporate role.
I also mentioned that you can download community skills. The most common place to find them is on GitHub. And the only pointer I want to give you is to be careful when you download community-based skills because what you're essentially doing is you're downloading external instructions that will fire off in your Claude Co-work, which can execute on your behalf, right? So, you will be at risk for what is referred to as prompt injections, meaning that you would have malicious prompts taking over your computer. So, if you're going to do external skills, then just make sure that you review whatever you're downloading carefully. You can pass the whole skill through Claude just to make sure that there's nothing crazy going on in there. But with that said, there are a lot of skills that you can recycle from the community which are really, really good. So, check that out in case you're curious about it.
So, by now, you know how to execute tasks using Claude Co-work, and the only thing missing is for Claude to trigger these tasks autonomously without your input. And this mechanism is something that we refer to as scheduled tasks, and you'll find the scheduled tasks in the left-side menu. So, if you click "Scheduled," you will end up in a view where you will have all your scheduled tasks summarized for you.
Now, if you click "New Task," you will have those two components, i.e., the execution and the autonomous trigger, available for you in one view. So, what you would do is you would set up a name for your scheduled task. You can add a description, and then you add the prompt. This is the execution that we have been working with up until now, and the trigger is set below. So, you can see that there is a tab here called "Frequency." So, you can set it to run at whatever cadence that you want.
Now, you can be creative here. You can go for something super simple like, "Brief me on the most important emails when I wake up." Or you can go for something more advanced like connecting your CRM to Claude Co-work and then have it review any changes that happened in the last 24 hours and give you a briefing on what your sales team have been up to so that you can support them accordingly. And you can build on top of that, right? If you see that your sales rep haven't reached out to a certain customer within a certain amount of time, then you can automate away an email and say, "Hey, is there a reason why you haven't reached out to customer X within the last 3 days?" So, you can stack this on top of each other as well, right? Really, the world is your oyster.
But let's do an example here. Let's say that we're going to fly somewhere. Let's do Spain. We like Spain. And we want to know when the right time is to buy the tickets. So, what we would do then is we'll add a name. Let's say, uh, "Flight Planner." In the description, we can say, "Recommends when I should buy my ticket to Spain." And here in the prompt, all we need to do is reference what we're trying to achieve. And if we're going to use an MCP, which I will do in this case, then we call that out. So, let's say, "I am flying from Stockholm to Barcelona, and I want to know when the right time is to buy the ticket. Use the kiwi.com MCP to evaluate ticket prices the last 3 months and use your knowledge of dynamic pricing to recommend me when I should buy the tickets." And let's set this up for a daily cadence. And this is using military time. I mean, here in Sweden, we use this by default. So, this is 9:00 a.m. And the last thing that is good to be aware of is that you can of course pick your model, but more importantly, you do your folder selection here. So, not that I'm expecting Claude to create any documents, but as a good practice, let's indeed select the folder anyways. And let's fire this away.
And now you'll see that it shows up as one of our scheduled tasks. And at any point in time, you can go into it. You can edit the same fields that we just worked with. And if you want to trigger it, you can run it right away, and you'll see what happens. And something that is pretty good to be aware of here is that you will have the history of when the scheduled task has run. So, if you want to go back to pick something up that you know that you saw a couple of days ago, you can just navigate over to your history field here and just select the entry that you want to take a look at. So, anytime you click it, you will actually be brought into the chat itself. It's starting to structure it up. It's telling me when the cheapest fares are: April 17th, May 15th. Obviously, I missed a very important step here, which is to tell it when I want to fly. So, now it gave me a couple of options. But hopefully, this will at least give you an indication of how this would work, as it is still giving us a recommendation of when the right time to buy is. And it's actually also recommending when we should be flying. So, again, creativity sets your limits here. Go wild.
As you work more with Claude Co-work, you will find things that you like and you will find things that you don't like. And in order to steer Claude Co-work down the happy path, you want to leverage personalization settings. Now, you do that by navigating down to your name and you click the "Settings" tab. And there are two personalization settings you need to be aware of.
So, first, you have your Claude personalization. This is not restricted to Co-work. This will apply across any of the three tabs that you have. And you can see in my case, I have a specific way that I want Claude to operate with me. You can see that I've structured it up under five different categories here. So, under communication, I don't want contractions like "I'll" or "we'll." I never write this stuff, right? So, I just picked up that it sometimes does this when it's writing for me. The second is the tone. So, "professional and direct, confident without hedging." You know, when it always kind of gives you this "buts" and "ifs." I basically wanted to take that out unless there are some true uncertainties that I need to be aware of. Then some inputs around the formatting. So, for example, "avoid excessive bold emphasis. Just keep it simple, right?" Um, in terms of problem-solving, I have a very particular way that I want it to approach problems because what happens otherwise is that I see how it solved the problem, and I know that I would have wanted it to do something different. So, rather than me doing that in every chat, I've just outlined the things that I wanted to do here. "Search online for any topic that is not obviously settled in knowledge." So, you've probably seen when Claude goes like, "Yeah, anti-gravity doesn't exist. That was just a prank." But in reality, it's just that the Claude model wasn't fed with information that is so recent that it, for example, knew what uh, Google anti-gravity is. I also have some steer around the sources. So, I prefer that it picks primary sources rather than aggregators. So, if I, for example, am asking about something around Anthropic documentation, then I wanted to actually pick up the Anthropic official documentation rather than third-party websites. So, feel free to copy this if you think this resonates with you. Uh, it's just a personal preference, right? But the idea is that you pick up on what you think works well for you and you plug it in here.
Now, Co-work actually has its specific section as well, and you'll see that for me, it's actually much shorter. So, if you just navigate over to Co-work and then go to "Global Instructions," you'll have this "Edit" button. And these instructions will apply to the Co-work tab only. And you'll see here that most of this is just sanitary things. There is definitely some echoing going on here in the first two points with top-down communication as well, but most of it is just the sanitary stuff. So, file formats. I prefer my file format the same way that we structure up the dates here in Sweden. I know this is not the way that you do it in the US, for example. You, I think you switch it around, if I'm not mistaken, but this is the way I do it, and then underscore descriptive name, and then I have some information around how I want to structure up the files, etc. So, the point is, instead of you having to get annoyed every time that it executes certain activities such as file structuring inconsistently across your different folders, then you can just plug that in here as a one and only structure that it will follow when it executes this.
And now you know how to use Claude Co-work. If Claude skills is something that you want to cover in more depth than what we did in this video, then I will point you to this video over here, which is where we'll take a proper deep dive on Claude skills. As always, thank you for trusting me with your time. And until next time, take care.