Loading summary
Transcript99 lines
- [00:00]
A
Hey, it's Mike Stelzner. What if you had an AI system that wrote your ad copy, generated the images and video, reviewed its own work and made revisions, all customized to your brand without you ever touching it? That's exactly what Caleb Cruise is teaching AI Business Society members to build on July 16th. No coding required. Join today at socialmediaexaminer.com AI and if you're listening to this in the future and you missed it, the full recording and resource kit are waiting for you. Visit socialmediaexaminer.com AI
- [00:43]
B
welcome to the AI Explored podcast, helping you put AI to work. And now, here's your host, Michael Stelzner. Hello, hello, hello.
- [00:55]
A
Thank you so much for joining me for the AI Explored podcast podcast brought to you by Social Media Examiner. I'm your host, Michael Stelzner, and this is the podcast for marketers, creators and business owners who want to know how to put AI to work. Hey, I'm just going to give you a heads up. Today is a longer episode because it is full of solid gold. This is one you might even want to listen to twice. Now, before we get into today's episode, I just want to let you know that I've been doing something kind of fun. I've been recording these short, less than three minute videos called Mike's Thoughts from the Porch and I'm publishing them on my Instagram, Facebook, LinkedIn and also on Social Media Examiner's YouTube channel. So if you want a little bit more from me about my thoughts on AI, my thoughts on marketing, and my thoughts on business, follow me on my personal profiles. I would love to hear from you. I try to respond to everybody if you respond, especially within the first 24 hours. So just find me on Instagram, Facebook, LinkedIn and also on Social Media Examiner's YouTube channel. With that, let's transition over to this week's interview. My special guest is an AI strategist and educator who helps businesses implement AI strategies. He's host of the Leveraging AI podcast. His course is Multi Agent Orchestration. Isar Matis, welcome back to the show. How you doing today?
- [02:11]
B
I'm doing awesome. I'm so excited to be here again. I enjoy every conversation you and I have. So I'm really excited about today.
- [02:17]
A
I'm really excited too, because Claude Cowork is like such a big thing that so many people want to understand. Before we get into it, what is one of the big misconceptions you see when it comes to AI Automation? Because obviously Cowork is moving towards automation and you've been using it for a very long time. And you're on the speaking surrogate, you have a podcast, you hear from a lot of people. What are some of the things that you hear from people that maybe isn't really true?
- [02:41]
B
I think the biggest thing is the fear factor, right? People think they need to be experts on something or need to know how to write code or need to be at the top of their game. And I teach this all the time, right? I do workshops on this and people come by. So how do I do this? I'm just ask Claude. Just instead of asking me whatever you wanted to ask me, just ask Claude and it will tell you exactly how to do it. And people go really? Like, yeah, 100%. Like, everything I've learned is a lot of experimentation and following a lot of people. But at the end of the day, the biggest misconception is that you need to be some kind of a guru, ninja, magician, you know, depending on who you ask, in order to actually benefit from this tool. And the reality is, all you need is to be able to explain what you want and it will walk you through the way to get to that outcome.
- [03:31]
A
Love it. Now, when it comes to cowork and automation, when it's done well because it can be done unwell, what are some of the benefits? What are some of the upsides, what are some of the unlocks that's waiting for people on the other side?
- [03:44]
B
I think the biggest thing is it becomes the thing that runs your business, right? It understands what is it that I'm trying to achieve across the board and understands my clients. It understands what I sell, it understands what I do, it understands my personal limitation and what my team can and cannot do, and it complements that in a completely seamless way. So from marketing to sales, to operations, to verifying that I'm following up on the things I need to follow up on on a personal level, on a management level, to creating content, to doing my finances, creating invoices, shipping. They like literally every aspect of the business it does while being fully integrated into my actual ecosystem. So it updates and talks to my CRM, my accounting system, my email, plat, my personal email, my Google Drive, et cetera, et cetera. Like every component of my tech stack, it is fully integrated with, and it can read from it and it can write back to it, and it becomes a part of the way the business runs.
- [04:50]
A
Love it. Okay, well, let's start by explaining what the difference is between Claude, cowork and cloud code. Let's talk about That a little bit. And Claude in general, just so everybody kind of understands kind of what the differences are. And then let's just talk about kind of the general concept of where we're going now with Cowork.
- [05:06]
B
So really, if you download the Claude desktop application, you have three different options. And they actually changed it last week in the middle of a workshop that I was doing. That was exciting. That's a whole other thing. But you have three different options. You have Claude, the regular cloud, like a chat, you have Claude code, and now inside the chat, there's two different options to switch between the regular chat and cowork. So let's first of all start with Claude code and Claude chat and what are the differences between them? So Claude chat is a chat. We all know that you chat, it chats back. It's a back and forth interaction where you're driving, you're telling it what to do every step of the way and you get answers that Claude knows how to provide. Similar to any other AI chat platform out there, Code was originally built to, like the name suggests, write code. And the crazy thing is, as much as it literally changed the world upside down, it was built as an internal tool, right? Initially Anthropic built this for their own usage, and then they figured out, well, maybe other people will want to use this. And then what happened is relatively quickly people started understanding that Claude Code can do a lot of other stuff because what it is, it's a agentic AI application. The fact that it's very good at writing code was the reason it got its name. And the reason it got its name is because they built it to write code with it. But very quickly people found out that it is extremely powerful at doing everything else. The downside of Claude code in the early days, now it's very different. Now it's an application and there's a user interface you had to run it through, like a cli, which stands for command line interface, which is basically like all dos. Like it's the furthest away from a proper user interface. It's a command line with simple text and you have to type in it and there's all these slash commands and it's like looks very, very techy and you need to run it in the terminal and people don't want to touch the terminal because they don't know what it will do to their computer. So there was a big pushback from the general population on using it, despite the fact it's very powerful. When Anthropic understood that, I said, hey, what if we take the incredible agent capabilities of Claude code and package it in a user friendly interface that A, people will find it easy to access and B, more importantly, people will be able to track what the agents are doing. So they build these interfaces that shows you what it connects to, which folder does it feed from, which files is it reading, what is the plan that it is following? So it's literally showing you a plan and it is crossing out steps as it is doing it. So you feel a lot more comfortable from, oh, it's not going to take over my computer and delete my bank account and disappear into, you know, the ether. I can see exactly what it's doing and it's a normal user interface. So now that changed basically everything because now you had tens of millions of people saying, hey, yes, I can use this, this is not. And they changed the name. Right? It's not called Code Anym, it's Cowork. I'm like, oh, it's co working with me. That makes a lot more sense. I don't want to write code, even though in the back end it does a lot of that without you knowing. But that's basically it. So that's kind of the story how that evolved. And that has taken off like no other product in history, I think. And now OpenAI is trying to copy it and Microsoft has taken literally Anthropic Cowork and embedded it into their Copilot environment. So now there's Copilot Cowork and now there's ChatGPT work, which will eventually probably do everything Anthropic does well.
- [08:41]
A
And it's really interesting because the way that humans interact with AI is evolving. Right. As you mentioned earlier, since ChatGPT, it's pretty much been you type in some text and you get a response and you type in some more text and you get a response. Right. And then Code and Cowork change that by actually doing things. Right. That's the difference. It's not just giving you text, it's doing things, it's going out, it's pulling information from websites, it's browsing the web. And because it's actually doing things, it's moving into this world of agentic. So talk to me a little bit about like where you see all this going with Cowork when it comes to the word agent or agentic.
- [09:22]
B
Yeah. So first of all, let's explain the three main differences between like an angelic environment versus just a regular chat. The first one is its ability to plan and execute on its own. So if you think about a chat, you tell it what to do, it finished. One step comes back to you need to tell it what to do again. You need to evaluate what it's doing, verify it actually did it correctly, and so on and so forth. In an agentic environment, you tell it what the goals are, what the limitations are, what it has access to as far as data and tools, and it will go and build its own plan. As I mentioned before, it will show you what the plan is if you're in cowork, and it will go and execute the plan step by step by step, so it can do significantly more sophisticated, complex processes on its own without you having to babysit it. So this is number one. Number two is it has memory, right? If you think about a regular chat, you start a new chat, it's a goldfish. It doesn't know what happened in the previous chat. And yes, I know these platforms now have memory. And it knows who you are and which company you work in, but not at the granular level of the actual work being done, the way all these tools have memory. And that's the cool thing. So I'll say something else I said in the beginning about people think you need to be a magician. People think that building agents and the whole concept of agents is something you need to hunt in a dark forest that has centaurs and unicorns, right? It's this magical creature that you go and find somewhere. And the reality is agents are text files. Everything in this whole agentic universe is files in a format called markdown. The extension is md and that's it. So it is set of instructions in simple text that tell it what to do. And this is how it has memory. It can write and read these files. So it can save what it did in the previous chat as whatever instructions, and then the next chat can read that, and then the 50th chat can also read that because the file is there and it tells it exactly what to do. So this is the second component is that it has memory, which is a huge, huge deal in the business, because it's. Everything in the business needs to know what happened, what is my brand guidelines? What did we do yesterday? We send this email. What kind of format was it in? How do we write proposals? All these things required to have a consistent kind of memory. And then the third component is the ability to use tools. As I mentioned in the beginning, my cloud environment connects to my entire tech stack. So I'll say something more extreme, beyond the fact that it can. It knows how to and when to. So it knows when to reach out to my emails and find an email. It knows when to reach to fathom and bring a summary of a transcript of a call that I had with a potential client. It knows when to update my CRM. So it doesn't just have access to these tools, it knows how and when to use them. So the combination of these things provide a package that enables it to act on your behalf, be an agent of you to do things in your universe wherever you give it access, which again, people are like, oh, want to give it access, then don't. That's perfectly fine. But you can control by giving it access to tools and different tools what it has access to and what it doesn't. So you can keep it contained to the things you have interest in it. Touching versus not.
- [12:40]
A
I love this and I don't think I've heard it said so simply. And I really, I'm going to reiterate what you said. What makes agentic AI powerful and cowork is this. By the way, we are making the declarative claim that with cowork you get all this. Number one is it can plan and execute on its own. And of course it can ask you for approvals along the way. We'll get into that, I'm sure. Secondly, it has the ability to remember what your preferences are. It learns from its interactions with you. And then the last thing is it can use tools. Right. This is what differentiates humans from apes. Right. Is our ability to use tools.
- [13:15]
B
Right.
- [13:16]
A
And probably apes have better memories than we do.
- [13:19]
B
Yes.
- [13:21]
A
But the truth of the matter is we're moving towards this world now where it's not, not just interaction, response, stimulus, response, but we're moving towards these abilities and these abilities now, when properly structured, can do very, very powerful things. So getting into cowork now, let's say that everybody who's listening wants to start with cowork. Obviously my understanding is you need to have a paid version of Claude. Is that correct?
- [13:46]
B
Yes.
- [13:46]
A
I don't think it's available for free.
- [13:48]
B
Correct.
- [13:48]
A
And Claude starts at $20 a month.
- [13:50]
B
Correct.
- [13:51]
A
If you're like, like Esar or me, you're probably on a max plan, which means you're paying for 100 to 200amonth. But you know, let's just start with like, how do we know something is a candidate for moving from just Claude projects and Claude skills and all these kinds of things into actually co work? So let's just start there because a lot of people are using AI and this is kind of a new way of thinking. Like, how do we know when we should start using CO work?
- [14:18]
B
Yesterday is the answer.
- [14:21]
A
But there are certain situations where it makes logical sense, right?
- [14:24]
B
No, 100%. And I think I want to touch on two things. One that you kind of mentioned that I think is very important, and the other one is the why. So the thing you kind of mentioned is like the $20 a month of 100 or 200. And people like you got to pay them $20 a month. I'm like, you pay $8 for coffee every morning. All it gives you is a kick of caffeine for the next 35 minutes. And you're going to do this again tomorrow. And maybe three times a day, potentially, for those $20, you are buying teams of people that can help you strategize, plan, execute, automate. Like all these different things. Like, from an ROI perspective, it is no offense to anybody, the dumbest question I've ever heard, literally. So putting that aside, pay whatever you need to pay to get as many tokens as you need in order to do whatever you need to do.
- [15:13]
A
And for the record, you can get a lot done on $20 a month.
- [15:16]
B
100%.
- [15:17]
A
Once you hit your limits, you can add another 20 bucks and then you can say, all right, I'm just going to go for the a hundred dollars, but keep going.
- [15:22]
B
And you do not need the latest and greatest fable 5. Like, you can use Sonnet and do incredible things.
- [15:29]
A
By the way, just so everybody knows, there's different models inside of Claude. You've got Haiku, which is their low level. You've got Sonnet, which is their mid level. And then you've got Opus, which is their high level. And then you've got Fable, which is their ultra high level. And I found that I use Opus for really creative tasks like Opus 4.6. But Sonnet 5 can handle almost everything el else, I think. Is that what you've all said?
- [15:50]
B
I agree.
- [15:50]
A
Yeah.
- [15:50]
B
I do a lot of my daily tasks with Sonnet. And then you get significantly more credits for the amount of dollars that you're paying.
- [15:57]
A
How do we know when it's time to move from just generally Claude using Claude as like we use Chat GPT, which generally people find as a massive improvement. How do we know when it's time to move from just Claude projects and skills to actually using Cowork?
- [16:10]
B
Yeah. So I think it's as soon as you understand that it can actually do real work for you versus give you answers. Right. And I think that is true for any business person for sure and probably every person on your personal life. Once you start planning projects and project could be your kid's birthday or planning a trip or anything that is more than just I'm going to ask questions and get answers. And what I wanted to say about the why. I want to give two very quick examples of things that I do with Cowork so people understand what's possible. I think that's the biggest problem that people have is like they don't know. So one thing that I have is all my content across the different social platforms is created by Claude. What Claude does is it actually goes and checks what the top people on the platform are doing in my niche. So the top AI performers on YouTube, on Instagram, on LinkedIn and so on, it learns from what they post regularly on what is working right now. And then it takes my content, My podcast, my YouTube channel, my weekly AI hangout, studies my community meetings, like all of that. And it finds stuff that I said actually me. So it doesn't just copy stuff out of the Internet because eventually people hire me. It finds stuff that I said that aligns with what is working right now. It is generating the content, it is generating the images, it's cutting my YouTube videos to shorter snippets and it provides it to me and my team for review so it can verify what it's doing. Makes sense. Once we approve or give it comments, it will fix whatever needs to be fixed and it gets posted it. This is example number one. Example number two, this is all with cowork. The second thing, second example, on the sales side, I get a lot of people who want to work with me. I have an onboarding call or a discovery call with a potential client. At the end of the call the AI verifies that there was actually a request for any kind of quote. If there is, it actually finds all the things that the client, potential client said. It researched them online, IT researchers their company online. It researches their industry to see what's going on right now. And then it writes a proposal better than the proposal that I can write. It lets me review it. I review it very quickly. I provide some comments or some changes, then it creates the final copy. It updates my CRM, it updates my Google Drive and it creates a draft email in the outbox together with the attached PDF. So instead of a 2 hour work, it's a 10 minute work x times a week. And so these are the kind of things so when you think about your day to day, whether you're in accounting or finance or leadership or strategy or consulting or marketing, whatever you're in. There are X number of things you do regularly that is a part of your work. And cowork can do multiple components of that, probably fully automated. I always verify everything through one way or another before it goes out. So this is the why, folks.
- [19:04]
A
I'm speaking to myself as much as I'm speaking to everyone who's listening. And once we get a tool and we figure out how to use it and it really helps us, which is absolutely the case for most people who are using projects and skills inside of claude, you kind of set yourself in a mental workflow where this is how I do it. I use claude, I give it projects, I develop skills. And it's really hard to kind of break out of that mold because it's been such an upgrade over the way we used to do things, right? And until you actually force yourself to say, okay, there's still a lot of manual things I'm doing. Copying things into a Google Doc, right? Out of an artifact, or I'm moving it from the Google Doc into a sauna, right? And I'm assigning it to someone else. Like, there's still all these little things that you do when you have people that work for you or contractors, and those are the things that are the next big leap of faith, effectively, right, Isar that we have to. We have to decide, okay, I actually want to automate. I don't want to just orchestrate, right? Because I feel like the human is the orchestrator and we're effectively giving that orchestrator registration over to AI. Is that true?
- [20:13]
B
100 perfectly said.
- [20:15]
A
Yeah. So, like, that's going to be a mental shift for a lot of people. It probably was for you too, right? You're probably using cloud all the time and like, all of a sudden you had to give it up to cowork, right? So where do you start? Like, how did you even, like, decide, okay, I'm going to try a little experiment here? Or do you have any thoughts on where we begin?
- [20:30]
B
Yeah. So I'll say two things. First of all, you kind of said, if you're already using CLAUDE and build skills and have projects, I think what you just said is probably 15 to 20% of the population. And I said that because I have people raise their hands when I ask this in workshops that I do, which happens every other week. So I get to see a lot of people in a lot of industries. Most people don't even have that. And you don't have to have that in order to go into cloud cowork and build agentic stuff. What I tell people is find the thing that you either do a lot and don't like doing, or that is just something that quote unquote, wastes you a lot of time, but it is a part of your role, right? Even if you like doing it, but you're saying, well, I can do better things with my time. And start with that. Start with one thing. Don't try to automate your entire business. Find the one thing that you either don't like doing or that cost you a lot of time and go to Claude and say, hey, open Claude, cowork session and say, here's my situation. I work in this and that company. I hold this in that role. This is something that I need to do every single day or every single week. Here's the data that comes into it and I'll opening parentheses for a second. If you don't have like an enterprise account, and definitely if you don't have a paid account, do not upload any business information into it. Closing parentheses. So give it the data that you have access to and say, this is what I'm trying to do. Here are the steps that I'm doing manually. Here's where the data comes from, where this is where it needs to go. How can we automate this process literally this way? Explain it as if you're talking to a consultant. And it will help you build the automation, it will help you build the skills, it will help you build the connectors, it will help you build everything you need in order to make this happen step by step, without knowing anything. Once you've done one simple thing, you're like, okay, this is not bad. How do I go to the next step? And how do I go to the next step and how do I go to the next step? And there are, are several different ways on how you go to the next steps. The first one, and I'll, I'll go back one thing before I go to the first one is you talked about the people who are using projects who are using skills. If you think about it, all these tools are basically projects that are connected together in a dynamic way, right? So what does a project have? A project has reference materials, knowledge files that it's connected to and a set of instructions. A skill is the same thing. It has connections to different files and references and a set of instructions. An agentic world just knows how to pool these skills as needed to complete a much more complex process. That's all There is a skill is already agentic, but it is a very simple, very specific kind of process. Very few things that we do have. Just one thing. Usually the thing that you do has four or five, six, 20 different steps. The agentic part of it is the fact that it knows how to call the right skills in the right sequence. And usually I'll use the same word that you use because it's the professional term. There's an orchestrator. There's one skill that knows what all the other skills do. And it will pass the data around and it will call them in the right time. It will move the information. So you said, how do you get started? And what I said in very simplistic way is defining the requirements. The way I look at it, and I'm going to talk philosophically for a second, but then we'll go back to the practical aspect of this very quickly. The world we used to know, and most people still do think that's the way the world operates. The thing that is slowing your company down from tripling, quadrupling is resources. Right? The anchor, if you want, that stops the ship from moving faster is resources. You need more people, you need more compute, you need more licenses, you need more stuff. That is no longer the case. You need three other things. Now, resources are good, don't get me wrong. Having more money and more people, more compute is great, but it's not a necessity. The three things that you need, the first one is you need great ideas. What is going to move the needle in my business tomorrow, next week? What are these things? So this is number one. Number two is you need judgment. What I mean by judgment and you know, like, I can define ugly, but I know it when I see it. Judgment is kind of the same thing. It's this, like, what is judgment? How do you define judgment? And so judgment applies in this case in two different scenarios. One is picking the right ideas very quickly. Once you, and especially if you're in an organization, you have employees and other people around you, you're going to have 237 ideas which are all good. You cannot implement 237 ideas, at least not at the same time. So you have to pick the ones that are the combination between easier to build and higher value to the business or to yourself on an individual, if it's an individual project. So you have to have judgment to figure out what these are going to be. The second part of judgment is during the work. And each and every one of you that has worked with AI on complex things we go back and forth. I asked Claude for something, it gives me an advice. I said, hey, wait a minute. But what about this? Said, oh, yeah, you're right. Let me check this. It comes back like, well, I don't think you're right. Go and do some additional research. It comes back, here's the research. Here's why I think this is better than what you think. Okay, now we agree. Let's move forward. Judgment is your ability to push back, ask additional questions and drive the boat. Right? So AI helps me. It's the engine. I don't actually do anything. The AI does the thing that needs to be done. But my judgment decides where to go and how to get there. So this is judgment. And the third thing which is going to bring us back to the tactics is writing really, really good requirements. Why? Because AI is stupidly good at execution. If you can define very detailed requirements of exactly what you want, what you need, how to get there, what the data is, what does it need to connect to, what is the flow. Like all these things. If you can define that in a granular way, AI will implement it in minutes. If you don't, you will wing it. You will make a lot of mistakes, it will cost you a lot more money and will take you a lot more time. So writing requirements is the number one thing from a tactical perspective. How do you write requirements? You ask Claude to interview you about the problem. I have a whole process that writes PRDs for me, product requirements documents. And it's not always products, but it's a process. Right? Here's what I'm trying to do. So I tell it the high level and then we have a process that is like 40 questions that takes me between 30 minutes to an hour to answer. But then Claude writes a 25 to a 40 page document that defines exactly what needs to be developed.
- [27:27]
A
So once we have this product requirement document or process requirement document, what do we do with that? I mean, I would imagine there needs to be human judgment on that. Right? And we need to look through that with a fine tooth comb, right?
- [27:38]
B
Yes and no. I'll be honest again, it writes these really long documents. I think I've read from one from beginning to end exactly zero times.
- [27:47]
A
So you're skimming them, basically.
- [27:48]
B
I know they're the right stuff because it's based on an interview with me for 45 minutes.
- [27:54]
A
Ah, okay, I like that.
- [27:56]
B
Right. So the content in there is everything I provided. I always ask it for a quick recap summary and usually it does it automatically. So it gives Me the document. But inside the regular chat you will say, here's a quick summary of what this is going to do. 1, 2, 3, 4, 5, 6. Like, okay, good. The next thing is let's create a development plan because this cannot be built in one day. It's not Rome, real quick.
- [28:21]
A
Before we get into the development plan, up until this point, technically you could be using Claude or Claude Cowork to build all this stuff, right?
- [28:28]
B
100%. Here's the reason why I do this in Claude Cowork, because I have a process that has four or five skills that does it with me. So the process that I developed runs inside of Claude Cowork because then it does beyond what Claude will just do out of the box. Because I build skills to make it better and better and better. Otherwise it's going to write you a page and a half of a document, which is still better than winging it. But it's not going to be a 25, 30 page document.
- [28:54]
A
Okay, so let's talk about this development plan. What is that exactly?
- [28:57]
B
So then I asked Claude, I said, okay, what are we going to develop first? My previous background, I used to be a CEO of software companies. In software companies, there's this concept of an MVP of a minimal viable product out of this whole 30 page document, what is the most critical thing we should start with that will be relatively easy to develop, that will give us a quick win without making it too complex. That's basically what an MVP is. And Claude is really good at speaking those things. So we'll give you, yeah, we should do this or this. And here's why I'm like, okay, so let's do one of those things.
- [29:27]
A
So wait a minute, we've got this requirements document. Are we just picking a piece out of there for the development plan? Is that what I'm hearing you say?
- [29:34]
B
Okay, let's take one of the examples that I give. You want something that will write your proposals for you, okay? And it needs to update your CRM and it needs to write the emails and it needs to update your Google Drive. And it's like it does 15 things. The one most important thing, the one, the core of this, that, by the way, when I started, I had this as a custom GPT and then I had it as a project in Claude and then I built the whole system around. Around it is taking a transcript and turning it into a written proposal.
- [30:00]
A
That's the core, the core Nucleus of whatever this product requirement document is. You would start with that first as part of your development Plan? Is that what I'm hearing? You say, yes.
- [30:09]
B
And then from there, once this is working, like, okay, now what is the next step? And I literally ask it in these words like, claude, where are we in the development and what do you suggest would be the next step and why? And you will tell me, we should do this, we should do this, or we have to do this first because otherwise we can do the other thing. Because the other thing depends on finishing this thing. Like, okay, then let's do the first thing first and then we can move to the other stuff.
- [30:32]
A
So how do we create a development plan then? Like, how's that different? Are we asking Claude Cowork to create the development plan?
- [30:38]
B
100%.
- [30:39]
A
Okay.
- [30:40]
B
And over there, again, this is where I apply my judgment. I'm like, no, I don't want this to be first. I think we should do this first. Or I don't like the approach you took. I think we should do it that way.
- [30:51]
A
Oh, so the development plan is the list of order of operations is really what I'm hearing you say. Got it. Okay, cool. So you start with the first things. What do you do? Then? Once you've got an order of operations,
- [31:02]
B
you just tell Claude to implement. It's really that simple. So if you think about the three main things that we talked about, writing the requirements, I'm asking Claude to interview me. So I tell it the general idea, but then I ask Claude to interview me again. If you just do this, just do this, ask Claude to ask you at least 10 questions about it, it will be much better than you just figure it out on your own. If you build a whole process, then it's going to be a lot more questions, going to be more granular, but it doesn't matter. So this is number one. Number two is it writes the PRD and then you literally give it back to Claude and say, hey, Claude, how should we develop this? So the PRD tells you what you're going to develop, it doesn't tell you in what order and doesn't tell you how. Which tools are you going to use? Which technologies are you going to use? Does it need a database? Does it need to be an external database? Can we just write an MD fil here locally on my computer and it will be fine. Like, this is the development plan that again, Claude creates, you judge, and you together start evolving and start developing. And the development is also. Claude said, okay, now that we agree on the first thing that needs to be developed, go develop it. And this might be a mix of four skills, an orchestrator, two NA10 processes to connect them all together. NA10 is like a third party tool that allows you to connect different pieces of software and a connector to your Gmail account. This is the first step and Claude will do all of this in one pass. Why? Because it has a very, very good requirements documents that tells exactly what it needs to do.
- [32:30]
A
You know what most marketers are doing manually? Ad creative, writing the copy, coordinating with the designer for the images, waiting on video edits, managing revisions, every campaign, every cycle. I wanted to fix that for our members. So on July 16th, I'm bringing Caleb Cruise to teach inside the AI Business Society. He's going to show members how to build personalized AI pipelines using Claude code that runs the entire creative process on its own. Copy, images, video, revisions, all customized to your brand. And you don't need to be a coder to do it. I'll be in the room for the entire session. I'm in every session. And I can tell you members are going to walk away with a clear, buildable system that they can put to work that very same week. If you can't make it live, the full recording and resource kit will be waiting for you. Transcripts, slides, key takeaways, the whole shebang. This is what we do every month inside the AI Business Society. Come and see for yourself. And don't miss this training. Join@SocialMediaExaminer.com AI again. SocialMediaExaminer.com when we were prepping for this, you said you developed this skill called Gotcha. Tell everybody kind of what that is and maybe some tips on how they could create their own.
- [33:49]
B
As you can understand, I build really sophisticated things with Claude. And when I say I build, Claude builds, but I come up with the ideas and the brainstorming and so on. In many cases, I don't have the perfect judgment. I think I do, but I know better. And so every time I'm on a major milestone, either a design milestone, a development milestone, a big decision I need to make, I have a skill called Gotcha. Gotcha is the ultimate devil's advocate. It will try to find holes and find mistakes that I've made across anything from the decisions to the tools to the complexity to the user interface, whatever the case may be, they will come back and say, say, here's where I disagree with you. Here's where there's gaps in what you said. Here is where common knowledge says otherwise. Once it finds those, it goes online and does research on each and every one of these things and comes back with best practices, other tools, suggestions how people like all these kind of things. And then it gives me a report, a very detailed report on what I should change in the plan before I go and execute it. And, and this saves me a lot of time and a lot of agony and probably a lot of money as well, because it finds things that I would have done wrong because my judgment wasn't perfect or just because I didn't think about something. And it is really easy to build. Like, literally all you need to do is to go to Claude and say, hey, I want to build a skill that is going to be the ultimate devil's advocate. I wanted to find all the loopholes, mistakes and gaps that I have in any plan that I have, regardless of what it is. And then I wanted to go and research those gaps online and come back with a set of recommendations on how to do it better. That's it.
- [35:31]
A
So when do you deploy this skill or is it part of your development plan?
- [35:37]
B
So it depends. There are places where it runs automatically based on stuff that I've created previously, but in most cases I will literally go and say, hey, I want Gotcha to take a look at this before we continue. That's it. Simple English. And then that reverse Gotcha. And Gotcha will do its thing and it will write its report. And again, I will get the executive summary, but I will not read the 12 page report that Gotcha created. I will give it back to Claude and say, hey, what do you think we should fix? Why do you think we should fix it? How do we fix it? What do you suggest we do next?
- [36:05]
A
Now let's get into actually using Claude Cowork, because up to this point we have laid some really important groundwork on the concept of coming up with a requirements document of whatever it is you want to do and then ultimately coming up with the development plan, which is the suggested course of action, and also developing a Gotcha skill. Let's actually dive into, like the things that everybody needs to understand about Cowork and kind of how it operates.
- [36:34]
B
At the end of the day, I think it's a lot simpler than people think. Going back to the biggest misconception you asked me about in the beginning, right? If you need the app, right? So you need the desktop app, you need the paid version, you click on Cowork instead of on Chat, and then you go and explain what is it that you want to develop. And once you do that and you ask it to ask you questions and interview about it so it has more information, want More context about what you do. This is step one. Once you have that and you developed whatever you want to develop, and sometimes the interview could be four questions and 10 minutes later you have the first thing up and running. The next thing is you iterate. I keep fixing and changing skills and processes and products and automations and applications that I built eight months ago, and I still update them. Why? Because A, the platform itself evolves and you get new requirements. B, you have different needs. C, you learn gaps from your current operation that you didn't think about it when you created it. And it's perfectly fine. Like I said, all you need to. And by the way, the update is the same thing. So how do you update the scale? I'm like, oh, based on what we learned right now I want to update the scale. Skill. That's it. Cloud will go in. I'm taking skill.
- [37:47]
A
So, Isar, I want to clarify specifically, there's the. Let's come up with the plan. And then I want to talk about, like the browser extension and APIs. I want to. I want people to understand how the heck it actually does what it does, because it needs access to things, obviously, and we've kind of skimmed over that. So I would love you to kind of explain how all that works, because that's the missing component.
- [38:09]
B
It. Great question. Thank you. And so a few things. 1. Claude Cowork. By default, when you open Claude Cowork, there's a button like a little folder icon with a plus on it. And you click on it, it will ask you to connect to a local folder. This becomes when I said it has memory and it can read and write files, this will be the place it read and writes files. My personal recommendation is I have one main folder that I literally call Claude Cowork. Try to say that four times, fast.
- [38:40]
A
It's a tongue twister.
- [38:43]
B
It's called Claude Cowork, and inside of it there's all the subfolders of the other things that I'm doing. And I'm always connecting the top folder for several different reasons. But the main reason is Claude understands the context of what we're going to work on. And it will find all the information, knows how to navigate, and it only
- [38:58]
A
has access to what's in that folder, nothing else on the computer. Yeah, correct.
- [39:02]
B
So it cannot take over your computer. The only thing it has access to is that folder. So by default, most of the work that Claude Cowork will do, same as Claude code, is work with a local folder and it will read and write files from over there, by default it will write MD files, those markdown files, just text files. But it knows how to create PowerPoint presentations and word documents and Excel files and whatever kind of file you can imagine, PDFs, like whatever format you ask it to create, it will create and you will save it into that one folder that you connect. So this is one source of universe that it has access to. But then like you said, you're like, okay, but I wanted to have access to the rest of the world. I wanted to have access to other tools and other things. So a few things. First of all, inside of Claude, if you go to Customize, which is a button on the left menu there you have several different things of skills. Plugins and connectors. What are connectors? Connectors are all these stuff that Anthropic themselves has worked with third party companies to build as approved connectors into the cloud environment. Most of the tools that you use, if using the common things that most of us use are already in there.
- [40:14]
A
For example, what are some of the tools that you're connected to?
- [40:16]
B
Google Drive, SharePoint, Notion, ClickUp, Asana, Monday.com, most marketing platforms, CRM, all these kind of things right now, if they don't exist, there are still ways to connect them. And guess what, if you don't know how to do this, you can say, hey Claude, I'm trying to create this in that thing. It has an API. So this is the second option, right? So tools that have API and every software out there today has an API. It does not have a formal approved connector to Claude, but it has an API. How do I connect it? And it will tell you what to do. Now in some cases there's a middle ground between these two things. That middle ground is called mcp.
- [41:00]
A
Yeah, I was curious about that because I'm seeing emails from all my software providers saying we have an MCP now.
- [41:05]
B
So what is an mcp? Almost two years ago, just over a year and a half ago, Anthropic figured out that everybody in the world will want to connect their tech stack, their CRM, their ERP, their marketing platform and so on into their AI environment. Regardless which AI environment, which means, let's take as an example HubSpot. Millions of companies literally have HubSpot. Every single one of them will either take their in house employees or hire external employees to develop an API connection to their AI platform so they can talk to one another. So that's really stupid. The work is going to be duplicated a million times. That's just for HubSpot. And then the same thing will happen for Salesforce, and the same thing will happen for Asana, and the thing will happen for Jira, and so on and so forth, so forth. So they said, what if we develop a standard interface that is a USB connector, meaning the equivalent of a USB connector on the back end. Somebody will need to develop the API once. And those of you who don't know what an API is, it's a language to talk to a software. It's the way one software talk to the other software. Just a protocol, communication protocol. So on one side, one time, somebody will develop the API to whatever tool. On the other side, there is a standard protocol that can connect seamlessly in two seconds to any AI solution, because it's going to be standardized. Again, my biggest analogy is a USB connector, right? On the other side of the usb, you can have a microphone, you can have a charger, you can have a camera, you can have a million things. But on this side, side, it's a USB connector. An MCP is a AI USB connector to any other software.
- [42:56]
A
And it goes bidirectional, right? It goes both ways, generally speaking.
- [42:59]
B
Correct?
- [42:59]
A
Okay, correct.
- [43:00]
B
So if you want the specifics, an API have functions in it. I want to read a document, I want to write a document, I want to find a document. These could be three different functions. So depending on which functions of the API were implemented as part of the mcp, because you don't have to implement everything thing, that's what the MCP capability is going to be. Now, the cool thing is, if you don't have an MCP for something, Claude knows how to develop it. So if you have the API access and you have the API documentation, Claude will know how to develop an MCP for you that you can now connect to every other place that you want.
- [43:37]
A
Okay, so just real quick here, because this is getting a little technical for people. In case your eyes are rolling backwards, people, let me translate. Okay, first of all, Claude has a lot of native connectors integrated directly inside of it. And there's a good chance that you would want to start there because those are the official ones that were developed in collaboration, presumably with the software companies. But these software Companies also have APIs that have existed forever. Claude, because it understands code, can go directly to the API. But now they're all coming out with these MCPs, which are kind of like a simple universal protocol that will allow not just Claude, but any system to work with them. And do you recommend letting Claude figure all this out, or do you recommend Just going with the MCPS or what do you recommend? Because it can get a little confusing, obviously.
- [44:23]
B
So I recommend the following. First of all, go with the standard ones that exist inside of Claude.
- [44:28]
A
Okay? Right.
- [44:28]
B
These are tested, approved by Anthropic themselves. Second level is if the third party company itself has created an mcp. It's also safe and probably tested and well documented. You now have two other options. One is going to third party vendors and there is a gazillion of them that develop MCPS for almost anything under the sun. Or you can develop it on your own. The advantage of taking a third party one that was created is that it's ready to go and you can just connect it and run with it. The disadvantage is, unless you're really good at going through this very thoroughly, you do not know what it is going to do with your data.
- [45:07]
A
Yeah, it might send it to some other party and you don't even realize it.
- [45:10]
B
You just give it access to A, your AI system, B, your whatever tool like your er, your erp, your CRM, your marketing platform and so on. And you don't know what it does with the data in the middle, you just don't. And so my recommendation is, do not use any random third party MCPs as tempted as it is. And you will find that if you go to Claude and say, okay, I have this third party tool, it does not have a formal connector, I do have an API connection to it. I want to implement an McP that does 1, 2, 3, 4, 5, 6. I don't need everything that the software does, that's just six things. CLAUDE will develop it for you. Now you know exactly what it's doing, you know where the data goes, you know exactly how to use it. And CLAUDE knows exactly how to use it because it's going to write a MD file, a markdown file that will explain exactly for Claude how to use the MCP it developed on own its.
- [46:02]
A
And all those credentials, those special security keys and stuff will live on your computer, right on your hard drive. So it's secure, it's not living up in the cloud, Correct? Right.
- [46:11]
B
And if you're on Mac, the easiest way to do this is to set them on keychain. CLAUDE knows how to go to keychain. It doesn't get access to the actual key. So even Claude itself doesn't have the actual security key. It asks keychain and say, hey keychain, I need you to connect me to this. And then it goes through that and then you get scanning disconnected. So there are a lot of tricky little ways that this data doesn't even really get saved on your computer in a file that anybody can read. Eventually if somebody steals your computer, like if you go to really extreme stuff, it will be saved in keychain in a secure way, all encrypted. And there's similar ways to do this, but if you're on a Mac, that's the easiest way to do that.
- [46:49]
A
There's another way, folks. That actually is the way that I started, which is probably the way that you started as well, which is the plush. Why don't you explain how that works a little bit?
- [46:58]
B
So let's go back and do a quick recap. So we said ways Claude can access your universe. We said Option number one is the formal connectors. Option number two is MCPS from the third party companies. Option number three is APIs directly to the software. Option number four is Claude knows how to use the Chrome browser. So literally anything you can access on the web, it can access on the web. And it literally opens the pages, it navigates, it opens menus, it clicks buttons, and it goes and does the thing on the actual software itself as if you were doing it, as if a human would do it. The only things it will not do, it will not fill in passwords. So every time it can, it just been told by anthropic not to. So it is not going to put in passwords, it is not going to copy any secret API keys and stuff like that. So there's a few things it will say, hey, I'm stuck on the login, can you log me in? Sure, you're going to log in. And then it will take over and continue and actually operate whatever it is it needs to operate. Now there are cases where this is extremely helpful and extremely easy because I don't have to actually connect to anything, I don't have to develop anything. And it will still do the work.
- [48:12]
A
Yeah, and it's cool because the way it works is it literally just kind of creates, opens a special tab, it works in the background, which is really cool. So you can still be working on your Chrome browser and it's working in its own little tab. And that's how I started working with it. So, okay, let's wrap this up in a little bow for people. This is an important little rabbit trail that we went down. So now that everybody understands all of this, how easy is it now once you've built all this stuff for this to do work for you?
- [48:40]
B
When I wake up in the morning about 27 things already happened. It is that easy. You know, that actually touches on in important point. You can schedule all the things you develop to run on whatever sequence you want. So there's things that gets triggered from the outside world. So if this happened, do this. But there are things that just run on a schedule. Every morning at 7am, every Monday at 4pm, every first of the month, you know, whatever, whatever you want, it will go and schedule and you will run these tasks so you can do things when you're not triggering them, when you're asleep, when you're on vacation, it will still go ahead and do all these things. Now, I'll say two caveats to that, and one of them is actually going away. One is, until four days ago, you needed your computer to be up and running and Claude to be up and running. Now they transferred all of that to the web. It's still clunky, it's still not working, but it will figure it out. So you can now have it run when your computer is not running, because it could run on the web. And there's a lot of other small ways on how to do that. The second thing that you can do, which is really, really cool, cool is you can work on it from your phone. So if you install the Claude app on your phone, you can now, until again, until a week ago, you had to use dispatch. And I'm not going to dive into this because it doesn't matter anymore because now you can have Claude cowork on your phone. So everything it knows how to do on your computer, it knows how to do on the phone as well. And you can now be on the road in your car and like, oh, I should have thought of that. And I don't want to forget before I get home because I will forget. And you just click on the button and you can talk to it with your voice. So you're still driving and you're safe and you will go and do the thing that you wanted to do.
- [50:21]
A
Folks, what we're talking about Here was released January 12th of this year, okay? This technology is literally only six months to seven months old. Okay? This is cutting edge tech and it is an absolute game changer. And I know it sounds really, really scary, but. But Isar does a lot of great stuff on his podcast and on his YouTube channel. So Isar, why don't you tell everyone where they can connect with you and discover more about you and if they want to work with you, how they would go about doing that.
- [50:51]
B
Yeah, well, you already mentioned the main thing you know the Leveraging AI podcast is the main place where I share very tactical things like we did here and in many cases more tactical. We actually share the screen and go step by step, kind of like a how to. And I interview the best practitioners I can find in the world to show show whoever listens and explain exactly how to do different AI use cases. There's the YouTube channel if you want to learn from me. And the biggest thing that I do is I teach workshops and courses. The workshops are for private companies. Any company that wants to hire me, I do either online or in person workshops. And for the courses, the latest course that has been selling like hot cookies is called the Multi Agent Orchestration course where I teach how to build some of the things that I just explained to you. And it goes step by step from the very beginning to people who know nothing. It is four sessions of two to two and a half hours each happening online. But with me live like you're not self paced, it's with me explaining all the different things and the people who are taking the course are building incredible stuff during and after the course, immediately after and for your listeners whoever will go and I'm going to give you the link so you can put it in your show notes so nobody has to worry about it. But if you also go to multiply but multiplied spell with AI instead of Y in the end so M U L T I P L a I dot A I you will find the courses and if you'll use the promo code AI explored, you're going to get $250 off the price of the course. The next course that will be available, hopefully not sold out yet, will be September. So that should give you a little bit of time.
- [52:27]
A
Isar Matis, thank you so much much for coming on the show and sharing your insights with us today.
- [52:31]
B
Thank you so much. This was awesome. I enjoy coming here and talking with you and the audience every single time.
- [52:36]
A
If you missed anything, we took all the notes for you over@social mediaexaminer.com A117 Be sure to follow the show on your favorite podcasting app and if you've been a listener for a while, we'd love a review. And by the way, you can also share this with your friends. You can tag me, you can find me. I'm Stelzner on on Facebook, Telzner on LinkedIn, Stelzner on Instagram and ikestelsner on X. And do check out our other show, the Social Media Marketing Podcast. This brings us to the end of the AI Explored Podcast. I'm your host, Michael Stelzner. I'll be back with you next week. I hope you make the best out of your day and may AI help you become more successful.
- [53:15]
B
The AI Explored Podcast is a production of Social Media Examiner.
- [53:22]
A
Before you go if you want to stop grinding through the AD creative process manually, the AI Business Society can help. Caleb Cruz is teaching members how to automate the entire process on July 16th. Full recordings and all the resources are included with your membership. Join at social media examiner.com AI again social media examiner.com AI.