# 142 Claude Cowork Masteclass for Beginners (full tutorial)

# 142 Claude Cowork Masteclass for Beginners (full tutorial) Speaker 2: What if you had four different AI employees all working at the same time, but working on different tasks? And then you had a fifth AI employee checking all their work, all from one simple interface. And that's actually exactly what Claude Cowork allows you to do. So in this episode, Mark Kashef is going to be walking us through Claude Cowork live. We're going to be doing a live demo auditing my website to come up with SEO ideas, design ideas, accessibility, content, All running in parallel with separate AI agents working at the same time. Now, Mark has been in this space for over a decade, so he's been doing it well before it was cool. And what stood out to me about this episode is he said himself, if you just spend a single weekend playing with Claude Cowork, you're going to be in the top 1% of people using it. So we're going to cover how to spin up parallel AI agents, when to use a quality assurance agent to check the work of the other agents, and the simplest way to get started, even if you're non-technical like me. So we're going to go ahead and dive into this week's episode. All right, Mark, why don't you tell the audience what it is that we're going to learn today? Speaker 1: You're going to learn about the beautiful world of Claude Cowork. Speaker 2: Awesome. I think Claude Cowork is an insane development in the AI space. And if you're not using it yet, you've got to. And if you have a Windows PC, unfortunately, you can't just yet. So if you're a Mac user, you need to jump on that. But before we jump in, Mark, why should people listen to you? Like, what's your background? Speaker 1: Yeah, I know. Great question. Sometimes I ask myself the same thing. But yeah, I've been in the AI space for 10 years. I have a Master's of Management in AI that I got in 2019, where we learned about GPT-2. I built like transformer models before. It was a cool thing. That's why I started my own agency in 2022 called Prompt Advisors, where we initially help people learn how to use prompt engineering in their business. That's scaled to now us working with hundreds of clients, hundreds of projects. And yeah, I've worked as a data scientist in the past at companies like Amazon and other startups. So I've been around the block a few times. So I might have something that the audience might want to be able to grasp. Speaker 2: I think for sure. And I love that you've been doing AI so, so far before it was cool. So I'm excited to jump in. Speaker 1: Awesome. Speaker 2: So the bulk of the episode today is Mark's going to be sharing his screen and walking us through Claude Cowork and showing us how that works and running through a couple of live examples. Before we get into that, let's just talk for a second or two about the difference between Claude Code and Claude Cowork. So Mark, for the layman who's not technical, do you mind just describing the difference between those two? Because they can be confusing. Speaker 1: For sure. So Claude Code is in a terminal and most people, most business owners, most non-nerds like myself will look at a terminal and run the opposite way as fast as possible. Up until the end of 2025, that was totally fair game. But now in 2026, the terminal has become non-technical friendly, meaning once you set it up, you can like interact with it just like you would ChatGPT except it's exponentially more powerful, more malleable and more flexible. But because people are still, you know, set in their ways, they've used Claude Cowork to make it that much easier on the eyes, that much easier on your soul, if you don't want to go and actually, you know, play around with the terminal. So Claude Cowork is one step above the Claude Cowork. Claude Cowork is as powerful as you can get. So you lose a bit on the horsepower, but you still get the experience with Claude Cowork. Speaker 2: And it's my understanding, correct me if I'm wrong, and you kind of alluded to this, but somebody asked Anthropic, who is the makers of Claude and Claude Code, I think they said, they asked somebody high up at the company, how long did it take you to build Claude Cowork? And like, how did you guys build it? And I'm pretty sure they said, we use Claude Code, basically they vibe coded Claude Cowork with Claude Code in like a week, I think, which is absolutely insane to me. Speaker 1: Yeah. So if anything, that's really proof in the pudding that Claude Code is the source and everything else is derivative. But for you, if you're non-technical, you can get the taste. It's basically like a gateway drug to getting to the promised land of milk and honey of Claude Code. Speaker 2: Right. Speaker 1: Yeah. So we're going to show that and I think it would help your audience if I actually show them what it looks like. Please share the screen. Speaker 2: And so for people listening on podcast recorders, we are going to kind of talk through what Mark's showing us on the screen. So right now he's just jumping into Claude Cowork and we're going to take it for a spin. I'm excited to see this thing in action because I have a Windows PC. Unfortunately, I don't have a Mac, so I can't jump into Cowork just yet, but I do use Claude Code in the terminal, which was a learning curve in and of itself. Speaker 1: My condolences on the PC. Speaker 2: Yeah, I know. That's what I've been told. Speaker 1: I'm joking. I've had one for like 20 years. This is a recent transformation. But when we hop into the Claude desktop app, we have three different tabs. We have one that's called Chat, and this is the OG chat that you'd be aware of if you've worked with Claude or ChatGPT. Then we have this new Claude Cowork, and then we have this dark zone of Claude Cowork, which you won't allude to or refer to today. Claude Cowork, like you can see here, it makes everything well categorized and compartmentalized. So you have a progress bar, you have a working folder where if there's any files in progress, you can monitor them there. And anything that Claude is using in context, you can monitor and see here. And if you need help with prompt engineering and you want to be able to control your computer or organize files, you can actually click on their pre-prompted buttons. It will add the prompts for you and all you have to do is just add the variables or switch the variables for whatever you want. So it's meant to really give you the steering wheel and more so the different guardrails you need to use the power of Claude Code without getting too deep in the weeds. Speaker 2: Got it. Now that's a great little preview there. So again, the difference between this and Claude Code is that you're just getting a little less of the, I guess, the capability using Cowork than you would like a Claude Code. Is that kind of the layman way to put it? Speaker 1: The actual sentence from the designer and creator of Claude Code is that they made Claude Code to be fully hackable, meaning you can make it whatever you want. You can make it do things it wasn't even programmed to do because at the end of the day, you're dealing with raw code. In a very code-friendly environment. Claude Cowork will naturally have some boundaries and restrictions and they're the ones who are listening to the market to see, okay, what feature do you want to see so don't overwhelm the non-technical user. Speaker 2: Right. Perfect. That makes a ton of sense. So then what are we going to be doing today inside Cowork? What are we going to be jumping into? Speaker 1: So I think that one of the most useful things that I could show your audience, especially if they're non-technical, is how you can use the concept of agents, real agents, not flimsy ones that you might've seen on YouTube where someone says, I just spun up this 108 node workflow with five agents that make me dinner and There's a lot of hype and BS out there. So the goal is how could I practically show your audience how to spin up agents for research purposes, which would give them enough to go deep and really play around with themselves. Speaker 2: Awesome. Let's do it. Speaker 1: Right on. So I need you for the prompt. So what are, what am I researching? What am I trying to get to the bottom of Corey? Speaker 2: Yeah, so and again, I'm going to selfishly use this as like a consulting session for my business. So our businesses return my time and we do AI tools assessments for small businesses. So we go out and find, basically we interview them, uncover their pain points and then prescribe Off-the-shelf AI tools that they can implement immediately to buy back 5 to 10 hours per week. So with that context in mind, we're trying to really beef up our website returnmytime.com from an SEO perspective. So what I'd love to do is use Cowork to either find keywords that we should be ranking for or find some tools that we should be using And we're here to help you in our SEO efforts. Does that make sense? Speaker 1: SEO, do you want anything on the design front, making the copy better, making the overall experience better or just the SEO? Speaker 2: Yeah, I'm perfectly open to seeing what a refreshed design would look like as well. I'm sure there's improvements we could be making from a design perspective for sure. Speaker 1: Gotcha. Okay, cool. So with that, I assume the website is at returnmytime.com? Speaker 2: Yep. Yep. Speaker 1: I'm not going to just spin up one task. We're going to start off with one task so I don't overwhelm your audience, but we are going to add multiple tasks in parallel and see if that works. So just to get started, I'm going to ask the following. I'm using a dictation software, by the way, just for those who don't know. So I'm going to be talking a lot, but I'm not talking to Corey. I'm talking to my other best friend, Claude. Speaker 2: Okay. Speaker 1: So I want you just to tell me what kind of agents you have access to off the cuff. So I haven't given you anything. I haven't told you what agents I want to spin up, but what do you come out with? You know, factory default settings. So you can do this to get acquainted with what you have out of the box. Cause sometimes if you want to build a planning agent, That already comes with Cloud Code. So this will actually break that down. It will tell you there's this Bash agent. If you don't know what Bash is, if you want to do any form of operations on your computer, Bash will help you. There's a general purpose versatile agent for researching complex questions, searching code and executing multi-step tasks. So this one is what you'll be using quite a bit if you use Cowork. You have Explore, which is more code oriented, so I won't focus on that. You have Plan, very helpful. And then if you want to ask general questions about Claude Code, there's an agent they've created that allows you to interview how this thing works. So if you want to go deep in the hood, kind of like someone reverse engineering how to build a desktop computer by breaking up the hardware, you can do the same thing. So this is what comes out of the box. And the cool thing is, I'll add also to your audience, is you can add what are called connectors, where you can enable this thing called Claude in Chrome, where I can make it actually not just go to the website and scrape your content, but go and visibly take a look at the website to audit it as well, which is why I asked that. So I'm going to first ask it, can you acquaint yourself We're going to start with this website and then let me paste this. I just want to make sure that Claude can see the webpage. Sometimes, depending on how the webpage is designed, it can't scrape it as easily. So right now, if we take a peek, it's actually opening it in a separate browser on my other screen and I'll bring it over here. It's going to ask me for permission to navigate your website. You can see right there, it's pulling up the website autonomously. It's going to scroll through and we don't have to watch it scroll through, but it will just get an idea of what the website is, go through and navigate it. We'll have all of this in context before we even go and execute the next different sequences. You can see right here, it's going through. And maybe we'll fast forward to when it's completed. Speaker 2: And I love that you're a WhisperFlow user as well. It's funny, when you said that you were going to dictate it, I was like, oh, I know Mark's using WhisperFlow. Anybody who's in this space is using it religiously. And I do as well. It's my most used tool by far, mainly because I probably type more than anything else. So it's natural that you would just speak. You're typing out loud. It just makes it so much faster and easier. Speaker 1: I would add one more thing since this looks like it's going to take a little bit more time. When you choose to dictate, usually prompt engineering was always constrained by your ability as the user to give enough detail and nuance to the AI for it to do what it needs to do. When you are yelling, swearing while you're talking to it, you can give all these micro instructions of everything you're unhappy with or everything you're optimizing for, which are infinitely more descriptive than if you tried to lazily type it yourself. Because when we type, we type the 80% and a lot of times we leave out the 20%. When you dictate it, you have less friction to not do that. Speaker 2: And I do find the one thing I struggle with sometimes when I'm dictating though, like a prompt, for example, is I find myself. Repeating myself a lot. So it's like, oh, I already said that earlier in my dictation. I don't necessarily need to say it again. And sometimes I contradict myself, right? I say one thing in paragraph one, and then I say the opposite in paragraph two. So there is, there certainly can be a downside, but overall, it makes it so much easier just to get your thoughts onto the page. And especially if you're doing a lot of prompting like you and I are doing, I mean, it's a complete game changer, saving a ton of time. Speaker 1: Yeah, the one little thing that you can do to avoid that pain is say, hey, I'm about to dictate my entire brain on this topic. I might have some things that are contradictory. Once you go through my mind sludge, come back and interview me questions to clarify anything that looks like it's conflicting or not. So in a way you create a purposeful and intentional speed bump. So it asks you questions and it gives you a second chance before it goes off and actually does something. Speaker 2: That's really smart. I actually never thought to do that. And as kind of like as a variation of that, you could say, hey, below is my brain dump. And before you process it, I want you to summarize it back to me in three to four sentences and make sure that I approve it. Right. So same same type of idea, because it might think that you're saying one thing when you're like, oh, actually, I meant this a little differently. And that could be the difference between a good output and a mediocre output. Speaker 1: Yeah. And the last thing I would say is sometimes if you don't know what you want, Usually, let's say you're having a bad day and someone asks you, oh, how are you doing, Corey? And you kind of say, like, I'm fine. And they have to, like, keep pulling information out of you. They have to have the initiative to pull it out of you. So if you can say, listen, I'm trying to do X. Ask me whatever questions you think that we need to get the full picture of how to execute X. And that would help you. It'd help you a lot. So, it's gone through, it's completed, it's full walkthrough, it's gotten a full picture of what you guys do. They got that little tagline of the five to 10 hours a week, ideal customer profile, resources you offer. So, this is all good to go and it asks me, what do you wanna do with it? Do you wanna brainstorm improvements, analyze competitors, work on content or something else? So, what I'll do now is we'll actually spin up a few different agents by just telling it that we want them. Okay, so you touched on it. I don't want you to do one thing alone. I actually want you to spin up a series of sub-agents. I want them to all work in parallel. One of them, I want you to go through all the copy that you've accumulated from the website and go and search for the latest and greatest documentation on AI SEO as of 2026 to see what this company can do to improve their SEO to rank better. The next thing is to have another agent to brainstorm general improvements to the site's layouts, the way everything's structured and anything that you think they're missing. Number three, I want you to take a look at the more so the color scheme. I want the third one to be a design agent where you go through, look at the hex codes, see the contrast they have and give anything in terms of recommendations on accessibility, on making things more engaging and easy to work with. And the last one is when it comes to content creation, if you think there are additional useful prompts or blogs they could offer, spin up one more agent to do that as well. So, I'll take my inhaler as this transcribes. Speaker 2: Oh, I'm excited to see this. I think this is going to be really good. Speaker 1: Yeah, fingers crossed. You never know with AI. And that's the key thing. You have to trial and error. Sometimes this works amazingly, sometimes less amazingly. So, when it confirms, I'm going to spin up four agents in parallel. What you want to look out for is you're going to see this main sentence here that says like a researching agent or something or working. And when you unravel that, that's where I like to spy and see what's happening. So okay, cool. Step one, we got the SEO analysis. I wanna see now, spinning up at the same time, the content evaluation. There we go. Site layout and UX improvements. Then we should have the last two. Color and accessibility analysis. This one we'll probably use the browser again to take a look at it. And the last one, Fingers crossed. There we go, content ideas. So now that all of these are initializing, now you're going to see them populate with different steps. As you see them populating with different steps, you can click right here. You could see that they're running all of these searches in parallel, if there are searches that are needed. Notice how the site layout and improvements doesn't need additional searches. Speaker 2: And so Mark, with Claude, like the difference between Claude and Claude Code is obviously, or not Claude Code, sorry, Claude Cowork. Cowork is spinning up agents to do all of these things simultaneously. Theoretically, you could do this in Claude itself, right? But it would have to be one action at a time. Is that right? Speaker 1: Exactly, so you can't spin up parallel agents at the same time in the normal Claude. Once in a while you can use a research agent, but I think one of the spoilers that I heard was eventually Cowork would actually replace Chat because it would essentially be irrelevant to have a less helpful mode and a slightly more helpful mode and then the full helpful mode right after. Speaker 2: Right, it's like why would you need Chat when you have Cowork, right? Or if you're using that already. Speaker 1: Why wouldn't you want agents is the bigger question. Right. Speaker 2: All right, so it looks like our little individual agents have gone out and done their thing. What are we looking at now? Speaker 1: Yes, sir. So one thing you always want to audit is what was done. So you can take a look at the eight steps and you can see it did Google AI SEO best practices for 2026 algorithm updates. It's gone through some guidelines. Schema markup for SEO best practices, keyword strategy, GEO. So that's all with us, by the way, not mentioning anything. It went and looked at local SEO, content marketing blog. So that's on that side. Site layout, there's only like one step because most of the context of the website, it already captured earlier in the conversation when it physically went on the website. So it didn't have to go through too much there. On the accessibility, it already has the website in context, but it went and searched the latest and greatest accessibility guidelines, colorblindness, everything that you'd need to make it more approachable. And then content ideas is where you had some searching. I would imagine it brought everything back. You can see we have a summary at the bottom. It says all four agents have completed their analysis. Let me give you the consolidated summary. SEO analysis, it walks through that and what the top priorities are. UX improvements, it gives you key findings. You have competing call to actions in the hero, which create decision paralysis, big gap between free action plan and 999 assessment, missing critical trust pages, top priorities, single primary CTA, add founder visibility. So Corey, you got to spawn more on that page. Restructure navigation. Oh, risk reversal, some form of guarantee. Then design, pure white on near black can cause halation. I don't even know what that is. Text bleed glow for 30 to 50% of people with astigmatism. Card to background contrast may fall below the three to one minimum for UI components. Specific recommendations, soften primary text. So this is giving you UI enhancement ideas as well as separation of the cards, which is cool because if you just gave this to normal Claude that didn't have access to this connector to go and search it and see it for itself, you would have to go and screenshot each and every page to then have it audited. The goal is doing the most with the least amount of input. Speaker 2: Right. Speaker 1: Then we have content ideas based on what you have and it comes up with those and then blog post ideas, podcast ideas. And then it asks me as a follow up, do you want me to compile any of these in the deliverable content? So just to your audience, I won't have to make us sit through this, but step one, I could say, cool, I want to be able to store this somewhere else. So can you create a Markdown file denoting everything that you found in a lot more detail? Right now, I know you're giving me the TLDR summary of it, but can you create a Markdown file and throw it into our either context or our working folder so that I can take advantage of this later? You can do this and it will create a series of files. We can go to the next step where what I like to do is one thing you can do in Claude Cowork is queue your next prompt. If you don't want to just have to wait and come back and you know what the next natural step is, you can do that. I can say, cool, can we spin up two more agents? One, to create a deck, a PowerPoint deck, nothing too fancy. It would be cool if you used our color scheme from the website and then created a deck on exactly what we'd need to do tactically and give examples. Right now, you didn't really give me examples on the SEO front, like what keywords have to change, what sentences have to be there. It should be a presentation that's comprehensive enough that I could read it, I could give it to my business partner and then they would be able to make the necessary decisions from there. And then I want you to spin up another agent that will act as the QA to look at the first draft of the PowerPoint to make sure the colors are good, the texts are good, the overlay and the distribution is good, and everything makes logical sense. So while I take another inhale, I can cue this up. And now you have two things ready to go at the same time. So this is running. This is running. And it's going to just keep going. Speaker 2: Well, that's such a smart idea to have a second agent basically to check the work of the first agent. I think that's something that most people would never think to do, but I mean, it makes sense just like you would in a real business. And that's where I think a lot of people go wrong when thinking about agents is an agent is essentially your employee, right? Like that's how you need to be looking at them. And if you had, let's say employee number one, putting together this big detailed PowerPoint deck, Of all these SEO tools and suggestions and best practices, wouldn't you naturally have another employee or maybe even you yourself look over it before presenting it to the stakeholder and making sure that it's correct? And so that's exactly what we're doing here. It's no different than how this process would work with two human employees. The only difference is we're using two AI employees. So that to me was a really interesting thing that you did there. And one other question too. So I know with like the initial prompt, we told it to do four different things, and we explicitly told it to spin up for individual sub agents, one for each of those tasks. My question is. If we didn't tell it explicitly, hey, use a sub-agent for this and then use a different agent for this, would it have done that automatically or would it have tried to do all of four of those tasks with one agent and therefore we probably would have gotten a less quality result? Speaker 1: Yeah, great question. So it really depends on Claude's mood that day and how good your prompt is. So it might, and I've seen it, Auto create agents on your behalf. If you can look at your prompt and it can actually parse out three independent or mutually exclusive tasks that could make sense to run in parallel. But if you prompt it in a way where it's kind of like gobbledygook and they're all somewhat interconnected, it might try to one shot it from the chat itself with zero agents. And where you lose out is not just on quality, but just to go back above while everything's running. Each one of these agents technically got a fresh chat of its own that you can't see on screen. So if your audience doesn't know, every language model has this denominator of how much context you can give it before things start to get into the hallucination nation zone, that I like to call it. So Claude, as of today, has 200,000 tokens. It's not necessarily one-to-one with words, but let's, for the sake of simplicity, call it words in the chat. Once you surpass 100,000, 120,000, that's where Claude starts to have amnesia of the beginning of the conversation. You start having to repeat yourself and not getting the results you want. So each one of these technically gets a fresh 200,000 context window of a separate chat where the entire chat is about that one task. So when you have agents and you use them like this, you zero in on focus versus doing a series of 80-20s across the board. Speaker 2: Right. Okay. That makes a ton of sense. And now have you ever given it tasks or worked on workflows where you get close to exhausting the full 200,000 token context window? Speaker 1: Yes. What it will do is it will do what's called auto-compacting. This actually didn't used to be the case. So up until December of 2025, You had to literally take your chat, summarize it elsewhere like Gemini and be like, can you make me a summary so I can continue a new chat? And then they added auto-compacting. So once it hits a certain threshold for them, that threshold is around 90%, which at that point, by the way, When it auto compacts to keep the conversation going, you don't know as the user what it has or has not included in that summary. Speaker 2: Right. Speaker 1: So you can still do it, but you don't know sometimes if it's continuing the context that you need to do what you need to do. Speaker 2: Right, because who's to say that when it summarized 90% of the conversation that it included all the key points that you needed, right? Maybe it kept six out of the seven key points and it's missing that one really important piece of context. So I could see that being an issue potentially. Speaker 1: Exactly, exactly. So if you go to the bottom here, it said it's created a report. And this is the Markdown file that I asked for. So this is a full comprehensive report on everything it did. And it stored it in a working folder. And it's called it return my time analysis dot Markdown file. And one thing I wanted to show your audience as well is we only had one piece of progress here. But let's say you didn't feel like spinning up agents and you wanted like four or five steps. One thing they added recently is if you see it going the wrong direction or you want to steer it back, you can actually intervene with ask a question or recommend a change for different steps that haven't happened yet. So it's cool because Cowork gives you a command center for either your agents or a particular chat. Speaker 2: So you're saying that like in that for that second task that we had queued up at one point that we can, while it's working on task one, it hasn't yet started task two, but we could go in and say, Hey, I know I gave you some instructions for task two, but let me clarify this point. Let me add something to it. That's what you're saying. Speaker 1: Exactly. So you get the best of both worlds. Speaker 2: Yeah, that's super slick because I mean, how annoying would it be if it's starting to work through that big second prompt that you gave it and you spent a lot of time putting that together and it was super detailed and then you're like, oh, now I got to start over again and it's got to basically rerun that prompt with this new context that they need to give it. But you're saying you can pretty much just inject that into the existing while it's already working. Speaker 1: Exactly. And you can also do the following where if I spin up a new chat and I say, you know what, I want you to work in this folder, my downloads folder. And you know what, I don't want you to use that Claude with Chrome thing. I just want you to scrape the website and say, redesign. This website and spin up a new version that's way more modern, pretty. I'm going to start with a titer copy and I paste this. It should come up with its own to-do list of how it's going to do that. First, it's going to fetch the page. Looks like it's failing to fetch the page, so it's going to try to find information about it. If needed, I'll let it do its thing. I'm trying to make this short and sweet. While this is running, we can still execute another task, which is go back to here and say, You forgot about my presentation deck about SEO. Where is it at? Speaker 2: Yeah, because what it gave us is like a full report, but I guess it didn't actually give us the report in slideshow format, which I mean, by the way, the report, just from the quick glance that we looked at, there's absolutely things from that report that we should be implementing. So that right there in the last, what, 15 to 20 minutes or so, And you gave us probably the next few weeks of work from an SEO perspective to get us significantly further along than we would even just working inside regular Claude. Speaker 1: Correct. And one cool thing is if it needs to clarify things, it has this thing called ask user input. So it'll give you a multiple choice of like, okay, what's the primary audience and purpose for the SEO presentation? And I'll say, uh, internal. And it says, okay, how many slides? Um, concise. You have steered it on the right direction and then you can go and let it do its thing. Here's the to-do list. It's coming up with like I alluded to. And then if you want to switch things up, you could. You could just click on this and comment on it. But now it should go and use what is called a skill, which I think is important for your audience to understand. Claude has injected capabilities for it to do that it can call on when needed. So if you want it to make an Excel file, if you want it to do your budgeting for you and you have this really rinky dink CSV where you manage your finances or your inventory or your products. It can create Excel files. It just has to know that. It will invoke the skill, basically reread the cheat sheet on how to make an Excel file and do it for you. So right now it's doing the same thing when it comes to how do I create a PowerPoint deck. So it read its skill. It realized it has to use this library called HTML, the PowerPoint. And now it's going to start the process of actually creating the first slide. Here's the first slide. And this is beautiful because you as the user can really keep track of what's happening and how. Speaker 2: And so you did a good job explaining skills there. And that's personally what I spend the most time doing is just thinking about what skills I need to create and actually going in and creating those skills and making them as detailed and comprehensive as possible. And how I've explained skills to people is that just think of them as A very detailed SOP or standard operating procedure for some tasks that you do in your business on a regular basis, right? And it can be anything. I mean, you can have hundreds or even thousands of skill files if you want, and they can be as granular as you'd like. So one example is I've got, and this is actually for my AI agent assistant. She has her own set of skills, but I have the same skills inside of Claude. But one skill that I've equipped her with is The ability to do podcast research for very specific types of podcasts that we then go and basically send them a cold pitch to get me to be a guest on the podcast. So she's got two skills associated with that. One of them is podcast research. So that's its own skill where she goes out and finds the podcast that we're looking for that meet our criteria. And then skill number two is actually crafting the pitch. So it's like podcast pitch writer where she then listens to the most recent episodes. Distills that into a very tailored, very personalized pitch based on the format that we use, and then it's ready to be sent. So a lot of times, at least the way I see it, and I assume you would agree, skills need to be like super specific. So like theoretically, I could have a skill that does both of those things, like one skill to do research and writing the pitch, but it just is a lot cleaner and I think a lot more effective to have One skill for research, one skill for writing, because at the end of the day, they are two separate tasks. Would you agree? Speaker 1: You did an incredible job articulating that. I feel like I'm not needed in this conversation. If I were to add on to that, when it comes to skills, you can think of normal life skills and how you can be amazing at each individual skill. So let's take the most mundane example. So let's say we're talking about brushing your teeth. Now, everyone in the world knows how to brush your teeth, at least hopefully. But what if you wanted to brush it in a way where you catch the most amount of plaque and it's the most effective way that doesn't hurt your gums and you want to be dentist approved, like top dentist approved. Well, it goes from being a general kind of thing you know how to do to then maybe creating a specialized skill where you watch like 10 YouTube videos of like the certified best way to brush your teeth. And that compressed knowledge and succinct, concise knowledge, along with maybe some diagrams of how exactly to brush up and down, that could be collapsed as a skill. And then you can invoke that skill whenever you're called on to brush your teeth. If we take that one small example, imagine that applied to every part of your life, budgeting, saving money, making money, driving, driving in general, All of those are skills and the best part of skills with Claude is you don't have to bloat its context window with giving it this whole user prompt or system prompt that says, this is how you drive, this is how you brush your teeth, this is how you make money. It calls on it what's called just in time. Meaning when it hears like there's something where it makes sense to use that skill, it will go read really quickly about that skill, exactly what's needed and go and execute it. And that's where it becomes that much more powerful. Speaker 2: Yeah, and that's very well said and that just in time is the key, right? Like it's not going to use the skill unless what you're asking it to do needs that skill, at which point it's going to go invoke that, like you said. So that's a great way to explain it. Now, so what are we looking at now? Is it done with what it was doing? Speaker 1: Great question. So it did finish quite a few slides. I think it's almost done. Yeah, it's really just creating the final one. I don't know if I want to reveal it before it's done completely, but we'll give a preview to your audience. So if you click on any of these files, by the way, you can always preview them because they live on your computer. So if we take a look at the first one, this is an example of like the first slide it put together before it went through and started organizing them all at the same time. It has all 12. It said now I've already read the documentation. I think it's trying to just finalize what the final deck should look like. And then it should give us a PPTX file. Now, while this is running, we could also show an example of a completed PowerPoint file, if that makes sense. Speaker 2: Yeah, let's take a look. Speaker 1: All right, so I just have to... Remember which, there we go. So this chat has one where I completed it. And you would be surprised, especially if you're a business owner that's using like 10 different AI tools for 10 different reasons. This PowerPoint skill is actually really, really solid. Like take a look at this. This was one shot. Speaker 2: Oh, that looks pretty good. Speaker 1: It is pretty, comes with stats. Speaker 2: I like those colors. Like I really like the colors and the font and just the design of this presentation. It looks really good. Speaker 1: And yeah, just to give your audience an idea, I gave zero Design instructions, zero structure examples. It came up with all this on its own. It spun up four different sub-agents to do research. It compiled that research. It created the deck. It had a QA agent go and look for overlapping text. And look at this. It's absolutely beautiful. Now, is it perfect? No. I want to also point that out too. You can see right here, if you squint, there's some overflow of text. This is where the QA agent could have been improved or I could have roasted the QA agent to tell it, no, look, you did a good job, but you missed this, this, this, go back. And here's the thing for your audience. Once it goes back, And it fixes it. I can say, cool. Now, after we've gone through this back and forth, create a solidified skill of how to QA PowerPoints and what to look for. Now that I've corrected you a couple of times on what you missed. So then if you do something more than twice, it deserves to be a skill and you want to do it the same way. Speaker 2: And what's cool about even just this one specific workflow right here of creating PowerPoints. I mean, there's probably three or four skills that we could chain together to create a really good PowerPoint. One would be like a specific PowerPoint design skill. So it's like, hey, we like all of our PowerPoints to be formatted and designed this way. Could be your colors, your branding, that could be part of the design skill. Skill number two could be And today I'm going to talk about the actual copywriting skill. So it's like, hey, skill number one was how do we design it? Skill number two is this is how we actually write copy for slides, right? Maybe we use bullet points, maybe we use numbers, maybe we're concise, maybe we're verbose. And then skill number three, like you said, could be a QA skill. This is how we check our work and make sure that we did it the right way. That's a three skill combination that would make your PowerPoints probably 10 times more effective than they are currently. Or you could just use something like Gamma, which I love. Speaker 1: I used to be huge on Gamma until I got this PowerPoint skill because it's one thing to use Gamma and live by their rules on how the PowerPoint should be made. But if I have a specific thing in mind and I finally get to it and I make it look like it's a big consulting firm style, then I just want to be able to crystallize that. So for me, tailoring is more important than just producing. Speaker 2: For sure. Yep. That makes a ton of sense, especially if you're doing a lot of these, right? And you want them to be really consistent. That's where it makes sense to invest the, you know, hour, two hours that it might take to put those skills together. Speaker 1: Yeah, and now we don't have to look at the output of this because I already gave your audience a little taste of the PowerPoint, but this is cool that it popped up so you could see. We now, after we made the presentation and are trying to make updates, we've bloated the context window. We've reached the logical end, so now it's compacting the conversation and then creating a TLDR to continue it. So this point is where you want to be careful, like you want to really handhold it from here on out to make sure it has exactly what it needs. Speaker 2: Right. Go ahead. I was going to say this is awesome. Yeah. Any other points you want to add to that? Speaker 1: No, it's more so like this is brand new, brand spanking new. So if your audience is watching this and you literally spend a weekend, you already will be in the top 10% of people that use Claude Cowork. And it looks like something they're really investing in to make better and better. So if you can just get used to this world and experiment, try to break things, see its limits and see what you can do with it, you're going to be so far ahead in the next few months. By just spending just a space full of time doing this and then once you're happy with it and you want even more power, the good news is it exists once you're ready to get there. Speaker 2: Right, which is jumping into Claude Code inside the terminal. And I use Cursor. That's where I'll use Claude Code inside Cursor. So that's really like the final frontier. And it's really not even as intimidating as you might think. I mean, I have no technical expertise, no coding background, no development skills. And it was very intimidating at first and I moved very slow and I still move pretty slow. Claude Code is incredibly powerful. And it's funny, Mark, you kind of alluded to the answer to the last question that I was going to ask you anyways, which was, so if somebody watched this and they're like, this is great, I want to use Cowork. But what is the easiest way for me to get my feet wet? You kind of already said it's like just spend a weekend messing around with it. You don't even need to have a specific goal in mind. But is there anything you would add to that for that person who's like, like, how do I get started in the simplest way possible? Speaker 1: Yeah, I mean, doing the most everyday task that would involve research. Research is the easiest gateway drug to get hooked on this. Let's say you're buying a brand new appliance. You're searching for plane tickets. Go and spin up four or five agents, each with a particular angle that they're looking for or a particular type of sites they're looking for and go and have them execute that research, come back to you, report on it, create a PDF guide and design of what they found. And just that one workflow, however mundane that is, would give you enough of a preview of what's possible. Speaker 2: A hundred percent. No, I think that's a really good plan. And Mark, thank you so much for taking the time to educate us today. I learned a ton and I'm sure the folks in the audience did as well. So with that said, Mark, where can people go follow you or go find out more about you or even work with you directly? Speaker 1: Yeah, first of all, an honor to be here and thank you for having me on the channel. I really appreciate it and I hope your audience actually benefited something. And in terms of YouTube, I'm also on YouTube. It's Mark underscore Kashef. I go very in the deep end, in the wild, but I also try to make things as comprehensible as possible when it comes to AI. And yeah, outside of that, I run a big community. So if you're ever interested in jumping into the very deep end of AI and really being in a community of hundreds of people doing the same thing, I run a community called Early AI Adopters and that's pretty much it. Speaker 2: Awesome. And of course, we're going to put links to Mark's YouTube channel and his community in the description if you're watching on YouTube or in the show notes if you're listening on Apple, Spotify, or any of the other podcast platforms. So Mark, thank you so much for your time and for everybody in the audience. We'll be back next Wednesday as always. Speaker 1: Cheers.