Founder's Guide to AI: Working with AI as Your Partner

Hello everyone, I hope you're having an awesome day. My name is Daniel, I'm a part of the Foundry Institute marketing team, and I'm super, super excited about our workshop series, uh, Founder's Guide to AI, with Trevor here. You can probably see Trevor on the screen as well. I'll introduce Trevor and let him take over, uh, in a minute. But yeah, this is the second time around we're doing this workshop series. The first time around we had over 3,000 people who signed up. It was a great success. A ton of people learned a lot of things. And yeah, we kind of decided to build on top of that because as Trevor will probably tell you, AI never stops. There's always things constantly happening, new software to learn, new skills to learn. That's a bit of a spoiler for what's to come. But anyways, before we really get into it, why don't you all introduce yourself in the chat? Let us know where you're joining in from, who you are. I know we have a ton of people from all across the world. We have people here from Australia, we have people here from Algeria, we have people here from the US, Canada, the UK. It's a pretty global audience, so that's always really, really exciting. Uh, so a bunch of really, really quick things before I let Trevor take over and introduce himself. We will be saving some time for the Q&A in the end. So if you do have questions, please drop them down in the Q&A section or in the chat, and we'll get to them in the end. If you do like what you're hearing, if you're enjoying everything, there is an emoji button at the bottom of your screen. Click on it and let us know how you feel, what you're thinking. And of course, this is just the first part in our workshop series. This is a three-parter. I will be dropping down the link for the next part in the series in the chat. Make sure that you don't miss that out. The idea here is for you to be able to take all the learnings from this workshop series and apply them as you would any skill that you would learn anywhere else. Without ado, Trevor, would you want to maybe introduce yourself, open things up?

Sure. Howdy, everybody. I'm Trevor. I'm the founder of CrowdTamers.com, marketing agency. I think we are one of the more AI-native marketing agencies that's not an actual AI agency. There's a ton that we do every day, that I do every day to build and improve and increase loops and, and what you can do with AI and what I have learned over the last 18 months of kind of bootstrapping myself into running an AI-native agency is what I want to share with you guys today. Now, there are two things I want to start with here. Uh, the very first thing, I think the most important thing for you to realize is how much things have changed and how much they have changed in ways that maybe You don't even know. And that big difference is the first thing to tackle here, right? And that is the, the way AI has showed up, the amount of velocity AI has brought into everything is easy to overlook. If you last checked out AI in December and you look at it again now, you almost wouldn't recognize what happened. When I originally proposed this workshop series, as a kind of a return. It was November of last year, and if you had asked me to make my slide deck for what it was I was going to share then, I don't think I'd have a single slide the same today. And so I had to do a bunch of stuff to make things ready for right now. You are supposed to see my Slack window, and I'll show you why in just a moment. The thing you have to realize is in order to make a good presentation today, I started the slide creation this morning. I wrote software just to be able to have the ability to present to you something instantly in a few moments with prep. And so that's what you see on the right half of my screen. On the left half of my screen, you'll see I'm asking you, what would you do if the cost of effort was free? And I don't think you realize how cheap it is to try a thing these days. So this I just opened up for the first time today. Idea Browser, Greg Eisenberg. He has a list of, I think it's like 2,000 different business ideas that are vetted, and any one of them has a market reach of more than a million dollars a year. And his idea of the day. This is not from today. Hang on, let me refresh this. This is today's. A call bot that waits on hold so you don't have to. What I'm going to do then is say, hey buddy, make me a Rails app that does this. I haven't changed on this one yet. Make me a Rails app that does this. And then I'm going to paste in— whoops, not my LinkedIn. I'm going to paste in this description. We'll just grab a bit of it here. Copy back to here and paste. And my AI is going to build this. While we do something else. I bring this up because this is all by itself a startup that Greg Eisenberg, who I assume is a smart guy and has validated, people would pay enough of them, $10, $15, $20 a month, whatever it is, to have an AI agent sit on the phone for them. And while I'm not even paying attention to it, My local AI agent running off of my computer is going to right through create the entire front end, the entire back end. I could set to hook up to Stripe and have all the payment links there as well. I'm not going to put the Stripe links live for this call because I don't want anybody here trying to pay something while I'm running this, but the whole thing can run. And while I'm running through the rest of my talk here, you can just watch and see what my AI agent is going to do to build this whole startup while we're talking. The total cost to me of generating this demo for you is about 7/10 of 1 cent. Something I could do while I go get my cup of coffee, take a break, come sit back down again in half an hour. The cost of delivering whatever it is you want to deliver right now has effectively vanished away to nothing. So what would you do if it didn't cost you to build things? Let me pop this up where you can see the updates show up in real time. There we go. So, that cheap delivery means the barrier between you and anything you wish to make real is you. And that's the first thing we have to really understand. Right? This idea, whatever it is, that we are building, isn't, like, the building, the execution has never been the problem. The problem has been for years now, but especially through today, the problem is How do people understand what's going on? How do people know why they should buy from you? How do people create and be a part of and make a story that's compelling enough that people wish to buy from you? And in order to be able to have the time and the attention and the focus and the skills to be out there and visible and building your moat around not what your features are, but your story, your brand, your positioning, your interests, your focus is— you have to free up time, you have to free up focus, you have to be able to take a lot of the work that used to be extremely time-intensive and begin to push it away. So there are Peter Yang published this, I think it was last week. I've mentioned this data is all as fresh as I can make it. He had the 5 stages of the AI native kind of journey. I've simplified things down a bit. I've made them less wordy, but the same core ideas are communicated here, right? Kind of the first stage of AI native fluency, what probably most of the people in this call right now are, is you're able to say, hey, I, uh, I will use AI. I will ask AI questions. I will use AI like a little bit smarter version of Google. And that's the, the first level where it's the thing you're used to typing into the browser window. And you'll get responses back. And then you might go, as I was talking to a buddy of mine today, the answers I get back half the time are insane, they're nonsense, they're garbage. So why would I ever do more with AI than this when all I get back is trash when I try and use the LLM for something? That's fair. But part of this is, this is a large language model. This is not a large maths model. This is not a large spreadsheets model. This is a model that's good at writing either in code or in whatever language you speak. And you have to structure the way you work with it around that. And once you begin to learn more about how to work with an LLM and have it engage back with you, then you can begin to trust it as something that is finding something you're doing and making it better. Now, here is where some folks go full seno. They go, I'm not going to make comments on LinkedIn anymore. I'm not going to make my posts on LinkedIn anymore. I'm going to optimize away. The thinking part of content creation. And that's why LinkedIn has seen a 64x increase in the number of posts put on LinkedIn in the last 18 months, because everybody has hooked up ChatGPT to their LinkedIn and is pumping out the same AI slop they've optimized, but they've optimized for the wrong thing. And that's very easy to do. It's an easy mistake to make, and probably everybody making their way up this journey makes that mistake. But then eventually, hopefully, you move beyond there. At the prototyper level, you begin to say, all right, I need a thing built. Maybe I need a Rails app to sit on hold for me, on call, because I don't want to waste my time waiting for my bank to get back to me. Okay, then you begin to trust and delegate and give away the problem of how would I even solve this. You're not yet saying this is something that lives in the real world. It's not something that is shipped and finished, but it's, you know, it's a piece of it. It's part of the solution. Then you might even say, all right, I'm going to trust you now as I am at this moment with, hey, you're going to build something. I'm going to validate it when it's finished. That something could be software. That something could be a proposal. That something could be a slide deck. Any number of things. You can trust it to build it. And finally, you start to say, I'm going to hand away to you this whole problem, something I don't understand. You're going to go solve for me. You're going to go take a big chunk of it, and then you're going to tell me how I should work with it. Those are the levels you work up. Let me just ask you, drop here in chat. If you were to cross these 5 levels, where would you put yourself? Are you down at level 1? Are you up at level 5? Are you somewhere in between? What is your feeling for where you are? I'm just curious to see. Wow, people up at agent, people up at builder. You guys are way up there. So then a lot of what I'm going to share to you, you might be like, yeah, I'm already doing this. Or maybe my spin, my take, my angle on this is a little different than the way you're doing it. I hope so anyway, and that you'll find the information coming in here is useful. So I'm seeing a lot of you guys a little further up, 3s and 4s. Not a lot of folks are saying, hey, I'm down to 1 and 2. Okay. So 1, that's heartening to hear, right? Founders need to be pushing themselves way up here. Right? If you're at researcher level, if you're still having everything being done where you work in the chat window and then the AI gives you an answer and then you go off and do something with that, you've really slowed things down. And a big part of what being a founder, and we're here at the Founder Institute because we are, is how much can I delegate and give away and moving that into speed. So, okay then, let's talk about what, like, what you want to get to move up from here. Now, most of you here have already said you're not here in the one errors. I'm not going to belabor this too much, right? Basic ChatGPT usage is where I find a ton of people, even founders, are. I know 3 different people who are founders. They're in the nonprofit space, so maybe it's a little slower. And this is the only use they have for any of these AI tools is they go, let me ask a question, maybe upload a file, have ChatGPT do stuff, give it back to me. And they're like, yeah, it's a little better. But then when I go, great, how do you move faster? How do you accelerate and go on to, I've defined a problem, let me fix it and go on to the next thing. They go, I don't even know how to begin to do that. It takes too much time. Okay. So daily work augmentation. There's a couple examples here of what I'm doing right now that even if you're all the way up at level 5 agent, right, building this level still has its value. Right now I do on an average day probably 5 calls. Some days I do as many as 12. Those days I'm not a fan of, but hey, you gotta do what you gotta do. These calls are mostly recorded and transcribed. And then in these calls, I say, yes, I'll do that. Yes, I'll do that. Yes, I'll do that. Will I do any of that? Probably not. I have team members who do a lot of it. And then also, I need to have on my days when I don't take calls some list of what I'm supposed to do. And I'll be honest, I do not want to read 18 to 24 call transcripts to then figure out what I was supposed to do today. I am AI loops set up, right? Go pull all of these transcripts. You know what my deliveries look like. I only have probably 20 things in my company that we do repeatedly. Scope for any of those. We know how much time they take. Book the time, put it in my calendar, away we go. Awesome, right? That is very much daily work augmentation. The AI isn't doing much except exercising a tiny piece of, we'll call it judgment, where it's told this is what work looks like. This is what scheduling looks like. Figure out how my life organized itself around these things. That's a bit of the agentic stuff, but also just a lot of help me plan. Right? I also have, every morning I post in my internal Slack channel to my company, hey, this is what I'm doing today. I don't post that anymore. The AI writes it for me. The AI writes it for my team. They get the stuff made so they can see based off of what has to be due now, what work must I do? Right, the project management role, a lot of it you can put into simple work augmentation loops. And then if you go, oh hey, that's wrong, I don't want to work on that today, right, you've got your check put in. Now there are a couple ways to do that. I'm going to talk about like what I'm building here, and you can see in some of my internal channels on products and things I'm working on, right, I use OpenClaw. That is a new frontier way to do agentic work. I also use other tools like Notice.ai, which is a lot like OpenClaw. It's been around a lot longer and it's very cloud-native ways to deliver these things. Whichever tool harness you use to enable these sorts of features, right? There's a lot of tech backend I'm going to go through and show you in a few moments. But making it work, hooking these things together is a time investment upfront. And when you put that time in upfront, it then shows up in way more time given later. Okay. Then you start to think about, I have ideas. There are things I want to build. How do I know if they're any good? Well, let me go grab— I've been posting a bunch of tests in this channel of the demo that I'm doing right now. Uh, let me see. I'm not even convinced this is still live. Let me see. Nope, I have already deleted a bunch of these off, uh, my, my regular schedule here, but I know I kept some of them live for the purposes of showing, uh, in— well, I mean, I can do this the hard way too. But I know I had an AI eulogy writer. Hang on. A-I-E-U-O-L-O-G-Y. Nope, I'm not going to find that. It'll help if I spell it right. You can tell I'm not used to using my keyboard anymore. I'll talk about that in a moment too. This is, I mean, For the purposes of my demo, I have— okay, I can do this, uh, here. If we do imagination quotient, imaginationquotient.com, is one, or scanzer.ai. Spell it right, Trevor. Scanzer.ai. There we go. These are Build validation, do it fast, right? This is a full tool that I built in half a day, third of a day, that will take any given GitHub repo. Do my security here. There we are. And, uh, here we go. Sign in. Looks at a repo and just says, hey, what kind of risk do these repos have from an enterprise information security architecture? Not so much a, is it going to hijack your computer and destroy it? And like, okay, Cool. I validate the idea. I've already got a couple folks trying it out. Maybe it's good. Maybe it doesn't. Built an absolutely insane website to see what I can do with AI design to make it do something other than the same purple and blue gradients that everybody has. And like, awesome, right? Seeing, seeing what you can push, what you can do. All of this is the builder validation. Prototyping at speed, trying to find what I can create and ship in a day or two and build from there. Then having conversations kind of like this one. Let me get out of the thread. Kind of like this one where we're shipping and launching a whole piece of software while I do something entirely different. And then as will inevitably happen, I'll go, hey, this is cool, but make this change. Hey, this is cool, but I want to see a difference, right? Scansr is a good example. Most recently, what was it? No, on Sunday, right? Hey, there's some, something that I want to see changed here. And conversational. I am not giving full product PRDs here necessarily, but rather I am running fast iteration loops where an idea comes to my mind. I drop it in and then I see what happens as a result. I make my changes. And then finally, things like, hey, I need to figure out a whole new business plan for something. How do I manage that? What do I do to make that happen? And creating a something like, and I do have here an entire structure around skills where I can say, and this is the downloadable you'll get for attending this, uh, on a bunch of different areas, a bunch of different questions, right? I can delegate and give to my AI an entire skill set. I'll show you how to do it in a moment of any of these things. Make it where Claude or any of the other tools I might connect to OpenClaw are able to grab this kind of information and use it to have a result for me. Those are the stages, right? The different ways to level up. And what you may find is as you try and graduate your way up here, you'll get stuck on a couple things. The first thing, and I'll go quickly on this one because most of us aren't here, is if you still treat chatbots Sorry, if you still treat AI like a chatbot, all you're looking for is an answer to a question. It locks you into that. If you can start to train yourself to understand how to give the AI what it needs to be freer, like a skill, like I just showed you, then you will be able to move up to context-defined prompts. So we talk a bit about Context. When you're running a big frontier model like Claude or Gemini or OpenAI, they have a context window. It's pretty big. 250,000 tokens is like 125,000 words, which is what most of the frontier models are at. Gemini claims to have a million token window. It doesn't really have that big a token window, but it is bigger than a 250K token window. So for those models, you can throw a bunch of stuff in and it will kind of sort all out in its head. When you're running stuff locally and you're paying on API usage for every token up and every token down, and by which model is using which type of token, it's going to be— you're going to want to avoid doing that because it's going to get very expensive to just dump a bunch of stuff in. And what you want to instead focus on is, how do I give my AI just enough knowledge to do the thing it needs to do? And especially if I'm running stuff locally on my computer, how do I put up a fence around it so that it doesn't decide, just for fun, let me leak my banking details to somebody who has access to, I don't know, my website chatbot, or delete my hard drive because it wanted to free up more space to store something that it thought was important right then. So what you need is to bring up— and let me do a new tab here— you need to get to where you are installing and running either Claude code in the terminal window like it's 150 years ago, other than when I was a child, or here in, uh, in Claude just natively, right? So we'll go here. All right, I'm running on Mac. It's going to be mostly the same for you when you're running on Windows. There's a couple things that are really important that you're going to want to have set up before you even try to do something like OpenClaw, which adds a bunch of other features into how everything is set up for your AI. At the basic level, what you want is to have either Claude Desktop or Claude Code where they are able to read and write to your hard drive. Because that's when context pruning becomes really effective, right? Once I can tell Claude, hey, go look in my meta skills, I don't know, go philosophy beginner mindset folder and help walk me through how I would approach solving this problem if I didn't know everything about this problem. Because that's the same idea as when a developer gets caught trying to write a piece of code and they say, let me get a rubber duck and talk to the rubber duck and try to explain to the rubber duck how to solve the problem. Because if I talk to it at that level, I might discover, oh, that's the mistake I wasn't seeing in my code. And so you can use the AI to do that or any other number of How do I do copy editing, right? Have me go through this, whatever the specific thing is. I can control that context when I'm running on my hard drive. I can't really do it if I'm just in the cloud. But if I just start a chat here and I say, I mean, I don't know, how do I write good copy? Good copy. Give me 5 bullet points. Man, I can't type anymore. Then what it's going to do is come up with generic stuff, right? The skill I've created that you'll be able to download here goes through very specifically how somebody who's the child of 2 English doctorates and somebody who's done copywriting for 30 years How do I approach solving this problem? You'll get a very different output here, but I can't easily tell the AI reference this if it's all in the cloud. On my hard drive, I can. I can say instead, uh, those look bad. Can you look up the— this is copywriting skill. I'm just going to grab this here. Copy path from system root. Now, because I can read my hard drive, I can say skill here and reframe, and then it's going to give me a completely different version of that based off of the skill I've told it. And now I can also tell it things like, write me— there we go— see, write me a landing page for What was the service we're building here again? Not that one over here in builds. Uh, this for this product, this product using, using these copywriting skills. There. And the output you get is tremendously better. Controlling that context window and telling it what it needs to know, really important. The other thing that's important for any of this context management is I have way up here at meta, actually here at Claude, I have a really simple guide on how to use all the information I have here. And what it says is, based off of the task you're trying to do, go look here for information about it. So that it knows here in meta, for example, it's going to go read here first and then have documentation on how to go tackle particular narrow things. Because I want to control that context load. Right? When I'm telling it to do work stuff, it doesn't need to think about me, my kids, my family calendar, any of that stuff. That's not necessary, and it costs me money. So either Claude Desktop or Claude Code, you're going to want to set up. And I will do this also. I will note as a caveat I've been doing this for long enough now where I am comfortable at the builder slash agentic level doing this because I already have stuff set up and containers configured to kind of limit what my Cloud Code instances can do. If you— have I spelled that right? Danger? I have not. Danger, O-U-S-L-Y. There's a flag literally called dangerously-skip-permissions. And what this does is make it so I don't have to have Claude ask me for approval on everything. Now, if you do this and you do this wrong, Claude might wipe your hard drive. Claude might, I don't know, insult your dog. Claude might do anything here you don't want it to do. So I don't recommend you start with dangerously skip permissions turned on until you have configured everything to make it relatively safe. The same is true here with Claude with desktop file access. I have a, uh, a connector which is Desktop Commander, let's read and write all the files locally. And I have Claude in Chrome, which means you can also open a browser and do anything. But again, I've got knowledge contained away from it to limit its capacity to just YOLO my banking stuff everywhere. Either of these ways, when you have these set up and you are reading and writing locally, you can control this context. Now in Claude, there is a specific— and let me go back here, hang on, uh, 80 skills folder, right? In Claude, there's a specific place where you should put the skills that I'm going to be offering all of you as a zip file to take with your download. And as you take Claude, and what you can do is install natively into Claude any of these folders by dragging and dropping them in. Depending on where you have it on your machine, they'll be in different places. But what you will want to do is have the skills there, have the routing skill there, and then have it set up so that it's going to look at development skill as you work through it all. So let me come back to this here. All right. Idea of the day, still building, I believe. Let me just check what we're doing here. Yep. It is built and hosted but not yet fully live. Okay, so back to this past the context level, right? When you move from chatbot to something that sits on my computer and you unlock this, what happens next? Then we have to trust, right? Trust, however, but verify. There's a couple of things, and you may have seen when you first joined I have an external drive there. Here looks just like this. I like how much the icon looks like the drive. It is not this drive. I wouldn't move a drive like this. It's actually under load. There is where I've got a totally different AI agent that runs completely locally. The Claude code that I am using and the Claude desktop that I am using don't have access to that drive. It is forbidden to it. So my banking info, my 1Password, my local copy of all of my emails that I'm using, uh, Apple Mail to download the files for— all of that sits on that drive so that you can't prompt inject my bot via an email, so that my bot can't accidentally destroy my credit by opening up 25 new credit cards or whatever or whatever or whatever. I have put a firewall in between what my really sensitive data is and what my merely business data is. And then with OpenClaw installed, I also have two different types of channel. I have an external channel of a client and an internal channel of a client. And my team is able to use the external channels to put client-facing stuff out. Clients can ask simple questions. On internal channels, we have access to all of our core documentation, all of our core loops, and team members can then ask for things like, hey, I need a new landing page shipped for this client. And we'll generate one and send it their way. So there's different levels of permission, even within the trust of, yes, I have an AI bot, but the external channels get one type of access and internal channels get different type of access, right? So all this building of fences is kind of like if you had an extremely intelligent, but entirely amoral, Super genius at your control, which I think is a pretty good way to think about what AI is. It is incredibly intelligent. It has no sense of right or wrong. A little bit of sense of right or wrong, but you can jailbreak almost all of it. And it lies to you with no bad feeling at all. So for things like, for example, trust but verify, we use AI to help generate client reporting. We generate 3, and then we look at the numbers, and if all 3 agree, awesome, there were no hallucinations. If 2 agree, then kind of like Minority Report from the movie in the early aughts now, you take the 2 that were in agreement and you assume that must be the true one because hallucinations almost never are duplicated. And so then we will use that as the basis for our actual reporting. Haven't gotten in trouble with that one yet. Maybe someday I will, right? You have to think about what would happen when you automate and give away the entirely remorselessness nature of computers, and then you have an AI involved in making decisions. So what we want to see, and And what we maybe have seen here, let me see if it says it's live yet. Oh, it is live there. Then we're gonna say publish to Railway so that it's going to be, live enough for us to poke around and see. The issue here, and I'm almost going to guarantee you, right, it already says, hey, it needs some things. It needs a Twilio account connected. It needs some thoughts about human logic. So it's going to publish a site where the front and the backend look really good. But because I didn't know today I'd have to have a Twilio CLI installed, I can't immediately hook up a Twilio account and have the actual call functioning. Yeah, within a few minutes, I'd be able to. I'm not gonna make you watch while I do it. So what I will do, and what you always have to do, is say, hey, it's really easy for me to build the app, right? What's the next step? Why are you opening this up? You have Railway installed. I don't know. The cheap nature of being wrong, of shipping an app, of saying, hey, I made Imagination Quotient, or I made Scansr, and maybe people like it and maybe people don't, means being wrong is effectively free, but being slow is what will kill you. So, okay. A big part of this is how do I get better and faster as a founder at offloading onto AI and trusting AI to deliver? There are a few things that I think are crucial to getting more native at this. One of them is the more comfortable you get at using a tool like WhisperFlow. And there's also, I mean, there's a bunch of them, right, that do really similar things. Any of them will work for your needs because when I, I mean, you can actually see, I didn't mean to click away, right? When I talk to WhisperFlow, I talk at 147 words a minute. When I try to type, you can see how awkwardly it goes, but I can say something like, sure, we can take care of the browser auth in just a moment. Previously you've deployed successfully via Railway, like yesterday. What changed? I can't type that fast, right? I also didn't tag my bot in the channel. So Let me paste that in. That level of speed and of fluency is one of the first things that if you are trying to make AI-native work part of it, I personally find this gives you a ton of velocity and depending on your personal preferences, you can have it read to you so you can do this while you're hands-free completely. I read a ton faster than I listen, so I always prefer to have the text pop up. I can easily fire back a response speaking instead of typing. That's just my highest throughput way, right? The change in just that, by doubling your typing speed, And then being able to, you can see here, right? It takes a while, even with this is, uh, currently I have Buddy hooked up to Gemini 3, right? Flash. Even with that, it takes a while to get the responses back. You could easily hit up 1, 2, 3, 4, 5 channels and say, hey, is anybody new? Hey, has anybody new signed up for Scanser in the last 48 hours? And just check. I mean, they haven't, but I can ask. And while one is thinking, another is still thinking as well. You can batch work out and then say, great, I can get back to the first one after I've asked the other three, right? By working at the speed you can think. And you start to delegate things away because I could, I could definitely go into the rails deployment dashboard for Scamzer and go look it up. I can also just say, you have access to everything you need, buddy, go figure it out for me. I would note this app, ImpressFlow, which is also scratch written, I put together over the last month or so. Right? Whenever I have a feature change or anything, I just go into AI, talk to it, get delivered faster that way. The thing you feed AI, right? The language you use, the files you put in, the skills you build yourself over time are the most important thing to begin to offload your work agentically. Right? All the skills I'm sharing with you are based on ones that make this agency go. They are around, and things I'll be talking about next session which is very specifically showing you how do you turn AI agency plus your story into the competitive advantage that actually makes you stand out because your features won't do it. So getting past the, the blocks of how do I ship faster on code or on proposals or on writing. How do I make something that is still uniquely me? Well, there's a bunch of skills that I've developed over the last 12 months that make that happen. But then you're going to have your own. And that's the other thing to think about, right? When you get— in a few moments, I'll have the link. When you get the link to your skills library that I've put together and you want to make use of any of these bits, you want to make use of these tools well, what you're going to need to do is make your own changes. You may discover that you really don't like the way I do copy editing. You think it's stupid. Edit it. Make it your own. You may discover that what's really important to you for proposals is this structure and then this structure and then this structure. And then do it, right? Any of these bits that you are able to customize to make it more like your business, you should, because that's how you actually unlock the right context that makes you, you. If we don't give AI context, we get the average of everybody on the internet, which is pretty bad. And all of us have seen all of the AI slop comments. You're absolutely right. Thumbs up emoji. It's not X, it's Y. You raise an interesting point. Any of these super common AI tells, they're not what you sound like. And with a tiny bit of work, you can help the AI produce stuff that is unique to you. When you give it that context window, then you get the output that you want to see. So if the product isn't the moat, what is? So Salesforce is a good example. It's a CRM. It's a, I think they first started developing Salesforce back in the 1400s or something. It's been a while now. And the legacy weight of it is tremendous and kind of crushing to anybody who tries to use it. So anybody could ship a little database that kind of strings together a couple of customer records here, maybe an email automation there. Like, if I can ship and publish a, an app from start to finish in an hour without even talking about it except the once. Anybody could ship Salesforce. So why do you still pay for Salesforce? Why do you still pay for DocuSign? Why do you still pay for Slack or Gemma AI or any of these things? You do it because they are solving specifically the things you want, specifically the ways you want them solved. And when the cost of effort drops to zero, I don't need— if I were to try to commercialize ImpressFlow, which I'm not— I don't need 10 or 20 or 100,000 users, all of them paying $15 a month to make the cost of ImpressFlow having been worth the development of it. I wrote a tool for me. If I ever commercialized it, okay, cool. It'd be, I don't know, 5 users paying $10 a month. Hee hee, there's $500 a month. I was never going to see. And that in terms of pure maintenance costs for me, I mean, I pay Cloud Code $100 a month. That's fair. I've got, I mean, I bought the AI domain. That's $250 a year. Beyond that, like, what's the operating cost of this? Right? Give me 100 users paying $5 a month or 1,000 users paying $5 a month, making money. Right? The moat isn't the feature set. The moat is who wants it. If somebody really wanted The capacity to get something where, you know, I can edit a Markdown file and instantly see the changes happen in the presentation. If that's something someone wants, cool. Do they want this particular brand of visibility into who the founder is and how it's built? And that's the story that makes it work. This is not the thing you need to invest hugely on the tech side. It's all about the story around it that makes it go. I gotta button back ahead here. I just realized I didn't enter a type in your number and go directly to that slide feature. So we're gonna go, oh no, we're gonna go back to that. The barrier then is your story. The barrier is who you are, and the better feedback you deliver into your AI is what makes it work. Let me go ahead, 4 more slides, then we'll go on to the next one. There we go. What's this mean for you? Well, there's two things I'd recommend you do. I'm gonna give you the easy mode to get started on building agentic AI. Anybody can do this. If you go with no knowledge of any kind of coding, of the terminal window, which looks scary to a lot of people. I'm old enough where that's how computers were back when I started. But if you're not that old, Where do you get started? You just go to Claude.ai. And it's going to say, cool, I can download Claude for desktop. A bunch of pop-ups here. Fine. There we go. I can just sign up and download Claude desktop. I think in order to make use of Claude desktop, you must be paying them at least $20 a month, I think. But for that $20 a month, you get something that it used to be you would pay hundreds of dollars, if not more, a month for. Because one of the things that Cloud Code— Cloud Desktop is going to unlock for you— sorry— is the complete amanuensis, the thing, the download I'm going to be sharing your way. It's going to have the ability to say, draft me an email, write me a landing page, ask me questions to help me firm up my business idea. All of these things that maybe you used to have a VA for, or an EA, an in-person executive assistant for, for $17 a month. It's pretty hard to beat. But also, once you've downloaded this and you've given it permission to access your local files, Then you can say something like, hey Claude, if I wanted you to download Claude Code to my computer, could you? Don't actually do it. I just want to know if you could. I already have it installed. If it tried again, it would cause me some troubles, but yes, it can. So Claude Desktop plus Desktop Commander extension. Oh, actually, could you just install the Desktop Commander extension for me, and then I'll get basically all the powers of Claude Code right here in Claude Desktop, right? And then yes, it could. We'll wait a moment for it to answer. Eh, Claude disagrees with me. Okay, regardless, I would also note Claude Code's available here if you click on it. In either case, right? Install Claude Desktop, pay the $17 a month, and then start to have it work off of your machine on the hard drive. That's the most important thing you can do today. But beyond that, right, take the downloadable you're going to get in just a moment from me and put it into your Claude skills directory and tell Claude, start to use these skills to do the work that I need done. I do recommend, if you can't set up Whisper Flow, there's also open source solutions you can install, or you can just use the voice-to-text on your keyboard. I find it is highly inaccurate, but maybe you'll have better luck. I just see now that my Rails app failed to build. That means there's something wrong with the execution of it. I can go try and figure it out, but it's beyond the scope of what we need to look at today in any case. And then I'm going to recommend you do is download the frameworks and try to apply them yourself. Now I could debug the Rails problem and fix it, but of course when you're live in an audience of hundreds of hundreds of people, that's when your automated build system Failed. I had it work half dozen times over the last week, but alas, not this time. In any case, grab this link, crowdtamers.gumroad.com/l/founder-skills. That you can go grab. I previously had folks say, hey, I can't access stuff off of your website, Trevor. Well, I'm fortunate enough to be one of the crowdfunded investors into Gumroad, and their platform is quite mature. So if my website wasn't doing its job last time, I know Gumroad will work. It'll link you directly to a zip file. Download the file, unzip it, and then ask Claude, where should I put this? I guarantee you Claude will put it in the right place for you. Now you have two other sessions coming up. This one was kind of the theory and the mind shift, right? You have to spend a good couple of hours figuring out how you're going to start to use this for yourself. There are direct applications for you that you need to use, but what they're going to look like specifically for you can be different for everybody. But the next two are very hands-on workshops. The first one is building your AI content engine. You need a moat, and your moat is not your features. Your moat is your you. And CrowdTamers, right now we have 20-odd clients who we explicitly build their moats for them and have tremendously high ROI doing so. CrowdTamer itself, I spend about $500 a month on my moat. I make about $20 grand a month back, and I'm doing it at the small scale because I only want to grow it so fast. I don't want to grow at crazy speeds. So that moat and how to build it is session 2. Session 3. Is going to be how do you distract signal from the noise and begin to take all of the data coming your way and make sense out of it. There's a lot going on and it's really hard to keep on top of all of it, but we'll help you make sense of it. Hopefully I've helped you see a bit about how to solve some problems. I've helped reframe some of your thinking, and hopefully something you found useful today. That is my time today. We've got about 15 minutes, I believe, Daniel, for Q&A and follow-up, and I'm happy to answer questions.

Awesome, awesome. Thank you so much, Trevor, for that presentation. I see that we have a ton of questions, a ton of appreciation as well. And what's super funny is while you were presenting on the topic of just how fast AI moves, uh, Anthropic literally just released another version of Sonnet. So now we have 4.6 out while you were presenting.

So there you go.

There you go.

That's what I, I, I have been using 4.5 like crazy. I would note when they pushed 4.6 for Opus, uh, 2 weeks ago or 3 weeks ago, they deprecated all of the endpoints for Opus instantly, and suddenly all of my workloads failed. And I was like, is, is Anthropic out? No, they just They just vibe-coded their way into launching their newest version of Opus. So Sonic 4.6 is live today. Given how much I like Opus 4.6, I bet it's going to do pretty good for y'all.

Yeah. So in case folks, like Trevor said, if you do want to get started with AI, I think Claude is a great starting point. So do check it out for yourself. It's always improving, so that's always nice. But I don't want to spend a ton of time there. Let's move into questions. So we have a question here from So Bodip, if I apologize if I got that wrong, but they ask, how are you measuring and monitoring if your AI agents are performing as expected?

So there's a couple different failback loops I have in place. As I already mentioned, when we use AI for client reporting, right, we generate 3 times and then we pick whichever number shows up at least twice. And that's a really easy Basically maths check with just Python after the report is generated. That one is very specific, right? I can measure the math output of a, uh, of a reporting table and say, was it delivered in this format? Do the numbers all agree? Cool. That's one way to do it. Uh, beyond that, when we do things like, uh, landing pages, for example, what I look for there are, there's some measurable things like how long will the page take to load? There's some immeasurable thing, which is, do I like the copy? And if I don't like the copy, why did it come out wrong? I think probably all of you have seen hundreds of the same purple gradient AI website. It's like, uh, the, the COVID spread of AI, uh, landing pages. It is just, they all make them that way. And 9 months ago or more now, I started to go, great, I've got to make something that is deliberately anti-that. And so if you download the skills that I shared the link for, and we'll also be sending in the email follow-up, uh, in there is a very opinionated design approach of if you build like this, it won't look like AI. You may not like how it built, but then you can make changes to it, right? And that's the other way I'm doing it. I check the outputs myself. And then if I don't find they do what I want, I'm going to go, okay, how do I give the right context window to improve them for next time?

Awesome. So just a couple of things to keep in mind, folks. We will be sending over all the incredible skills that Trevor is talking about in our follow-up email, as well as a recording to this presentation. Remember, this is a three-parter, so we do have the next workshop in the series coming up. So I have shared a link to it if you directly want to sign up to that. So do check that out in the chat. Okay, so moving on to the next question. John here asks, how can you speak about tokenization? How do you avoid overspending?

That's a hard one. Uh, there's a couple ways to do it. One is just start with cheaper models. Like, I have OpenClaw, which was originally like, use it with Claude. That was the name, was originally ClaudeBot. And they got DMCA'd. And then of course, ironically, OpenAI bought them. I'm sure that the, uh, I saw the founder talking about why he never thought about selling to Anthropic was because the only love letters he got from them were from the legal team. And, uh, so okay, like, this was designed to be used with Anthropic, it's expensive. Claude's API is the most expensive out of all the frontier models. So then I've done a couple different things. One is I have a Mac Mini M4 Pro with 64GB of RAM. I know everybody bought a Mac for, or Mac Mini for Open Claude. I did it back in December before Open Claude was a thing. I have to let you know I'm not that much of an impulse buyer, but I did do it because I had a very specific need. I need to generate about 15 million words of text for a particular AI training program I am doing. And at API costs, that is half the price of a Mac Mini. And once you go, okay, well, I can buy a Mac Mini for basically what I foresee my usage workload is in 2026. Might as well do it. I'm running local models, Quen 330B MOE. That just trips off the tongue nicely, doesn't it? Uh, Quen Encoder 332B, I think, is the other one. Uh, and these are local models. The only cost is electricity. They do have— they're smaller, they're dumber, they're slower. Part of why the delivery of the app that I was trying to launch live on Rails right now failed is Quen 3 coder doesn't handle fail loops well. If I was using Opus, it would have handled it all, I'm sure, and it would be deployed. But I was running it locally on my machine because that's what costs 7/10 of 1 cent to do, is just the electricity bill, right? Um, the other thing then is you can set up, if you use OpenClaw, a, uh, a router where based off the complexity of the thing you're asking It's going to route it to different models. And so if you're asking really easy things, you point that to your local model. If you're asking really hard things, then you go to KIMI 2.5 or StepFlash 3, or I mean GLM, or like whatever, any other model. And what I would suggest in general is get your build where you're happy with it and the expense isn't crippling, and then leave it because it's really easy. To keep tweaking and optimizing and doing this and that and the other thing. And it's busy, but it is not productive.

So just, just to follow up on that, and this kind of came up in the chat as well, you spoke about PRDs and then not exactly PRDs, but how you prompt when you build. So you're essentially like prompt and quickly iterating. So yeah, If you could do like, maybe a PRD versus prompt, just like how you do it. And maybe if you could, for the folks who don't know, if you could explain what a PRD is.

Sure. PRD is Product Requirements Doc. I've also seen them called PDD, Product Design Docs, TDD, Technical Design Docs. Find any 3 letters you like. It is immaterial. Now, part of what happens when I am building is I don't have to give a bunch of PRDs because my skills contain them. Right? My skills describe what good copywriting for a product is. My skills describe what good design for a product is. So I don't have to put those into the product requirements doc. My skills already lay out my expectations there. So it allows you to basically assemble in the AI's mind a PRD at runtime where I can say something like, I'll go do a new channel here. Um, these usually ship faster. Uh, I'll go this one. Yes. And I can say Hey, Buddy, please make a landing page for crowdtamer.com where we talk about, uh, we're the only marketing agency run by a sentient mustache. And then I have to tag Buddy in. Whoops. Oh no. I have to tag Buddy in because I have it set where it only listens when I tag it. And then it will go build this, but it already knows the language I want to use and the designs I want to use. So over the course of the next 3 to 5 minutes, as that builds, it's going to follow the existing PRDs that are there. So this is a terrible PRD in general, but when you give it all the context manipulation I'm talking about, it helps. I'll let this run for a few moments. What's the next question?

So you did mention OpenClaw and putting it together and how you're using it. So my question would be, what is it that you can't exactly do with OpenClaw or any agent, right? And what would you like reserve for like the founder to do themselves?

So I use OpenClaw and one of their competitors called Notice AI kind of in parallel. And it's, I mean, I'm probably a nut job for separating things this way, but whatever. What I have found through using these really extensively over the last 6 weeks, and that's really, I mean, OpenClaw became a thing end of December, so. There's not like there's that long to be fluent in OpenClaw. Notice AI I use for cloud-native workloads, for things like, hey, I've got this email that I want to automatically send replies to. It lives in the cloud. It does those workloads. Unlike OpenClaw, it doesn't require a bunch of fiddling and configuring. What language model do they use for Notis? I don't even know. I don't care. Right? I pay Notis whatever flat fee a month and I hope they make money. Not my concern. OpenClaw runs on my API endpoints. I pay for the usage. And last month I was paying, I think, like $300 a week of API credits. That was painful. And I wasn't even using Claude. I was using Kibbe 2.5. So I went, okay, how do I run more stuff locally? So this, this landing page generation should be running locally and hitting up Quencoder 3 instead of running all the way to the cloud. Then it will eventually prepare something, push it to GitHub, and then make the GitHub page live on Vercel. Cool. That's the workflow here, right? It's a very discrete thing. Write a page, save a page to this particular repo. And then push it to the repo so that Vercel, which is orchestrating stuff on the other side, makes it live, right? That's something you can give away because it is very discreet. If I have a thing like, uh, I've got a customer who's angry, I can't give that away, right? I need to be the human who determines these things. If I've got a thing like I want to make thought leadership, right, Right? My thoughts are unique. They're the part that makes me, me. If you try and push thought leadership content onto AI, you get crap. If you try and put most any human effort into AI, you get crap. You can do technical documentation. You can do landing page creation. You can do a whole bunch of knowledge work, but stuff that is, if it's something you could sensibly give to an intern, that's actually, that's what Notus calls themselves. Come to think of it, they're your AI intern, right? If you can give it to an intern, then AI should take it. If you couldn't trust an intern with it, because I don't know how many of you have ever used interns, interns will do anything with no thought for the consequences frequently. Because this is their first job and they don't have guardrails yet. Just like AI, you give it to an intern, you can give it AI. If you wouldn't give it to an intern, maybe you should still be doing it.

And on the point of like, you know, making sure that things don't go wrong, right? You did mention you do use quad code, right? And I think one of the concerns that people have, including Subodh in our chat, was how do you protect against Claude code you doing like irreparable damage to your computer?

Sure. So there's two ways to do that, right? One is, and I already showed you, I mentioned, right, I have a hard drive where all the sensitive data sits. I would also note that I have— Mac has, um, Time Warp. Time Warp. And those backups push to that hard drive too. So if somehow Claude bricked my machine, I have, I think, either every 12 or every 24 hours an image of the whole machine saved. So worst comes to worst, like any disaster prevention system, I have a backup. The other thing is I'm running Claude in a Docker container. A container is basically a virtual computer sitting inside of your actual computer, and that virtual computer, you can say you only get access to this stuff. And then even if Claude goes insane and wipes the hard drive. It only wipes the virtual computer's hard drive. Now, I would note, I don't have the least idea how to build a Docker container. It is not something that I am able to do, but I can tell Claude code, build a Docker container, and then run in the Docker container next time. Show me how to build a virtual environment. Actually, don't even show me. Build one, and then show me how to run you from it next time. All of the skills and kind of pulling myself up bootstrapping I've been doing is mostly just learning what to ask the AI. If you sat me down on a computer terminal with no AI assistance, I would fall back on my development skills from 25 years ago, which were so bad I decided to never be a developer again.

So we have a bit of a good follow-up to that here from John, and John asks, uh, do you have any concerns about moving too quickly when it comes as a team and not just as a founder? And then also inviting, you know, there's like security risk and also like, uh, operational risks as well. How do you mitigate those as a founder?

So I have an advantage. My team is small. There's 4 of us, right? We can operate in a fairly high trust environment. I can say things like, hey, in theory, I am pretty sure. Yeah, because, uh, my internal AI agent that is connected to all of my internal Slack channels, because that is has access to my Google Drive. In theory, my team members could read anything off of my Google Drive. I'm not convinced there are sensitive things there, but there probably are just because I don't even know what all is stored there, right? They could, in theory, create tasks in Notion as me. And I mean, I don't know why. They could, right? Would they? I know my team well enough to know, no. One, they're busy. They have their own life to live. Two, they are people who I've worked with for multiple years. One of my team members next week has her 5th anniversary working at CrowdTamers. Like, these are people who we, I have that level of trust with, so I'm comfortable doing it. If you employ 500 people, don't do that. If you employ 500 people, set up strong permissions and limitations and make sure that you've got guardrails up. The other thing is, if you've got 500 employees, and you say, here's an API endpoint, good luck. No, man, they don't wanna. If you've got 500 employees, what you do is you say, hey, everybody, You don't have to do a daily standup meeting anymore. All you have to do is the AI is going to call you every morning at 10 AM your time and just ask you, hey, I see this is what you're supposed to be doing today. Are you tackling it or something else more important come up? And you're going to answer them and hang up. That's their AI experience, but suddenly Everybody in the org gets transparency on who's working on what, what deliverables are on time, what's late. Like, there's a whole backend that you make happen. The endpoint that you expose isn't, here's an unlimited tool you can do anything with. And as a founder, when you sit down to use AI for the first time, kind of a similar thing, right? I could do anything. So what do I do? Oh no. Well, find a small thing and then ask Claude, how do I do the small thing? And it'll say, here are ways to do the small thing. And you go, cool, let's try one. And then you don't try to embrace everything. You just embrace the little bit that starts to give you momentum, and then the next, and the next, and the next.

And something that I found super interesting with using AI, especially like cloud code, right, where you're actually building things, is that you're learning a bit of the technicality behind building stuff, right? So it's not just you, it's like, you know, token in, token out. There's also like this process by which you as a founder, as somebody who's building stuff, someone who's working on a craft, you're getting better and better. And I think like that is probably a really, really important thing. And it's also a good opportunity, right, for anyone to pick up an AI tool and be able to teach themselves stuff.

I agree. And I would— a year ago, I didn't even know what a Docker container or a virtual environment was. I wasn't aware of how to set up an auto deploy hook to Vercel from GitHub. I'd never really paid attention to the existence of Ruby on Rails and Railway for auto deploys, like all of this technical stuff. I'm a marketer. I write words, dude. Like that was my lane. And you can blend them now, right? Am I as good as a real dev? Heck no. And a real dev will smoke me on, you know, why, why are all my things failing right now? I'm sharing my screen. I assume because my ad's trying to make me look bad, but a real dev would know why the pipeline's broken and it wouldn't have gotten wrong the first time, right? But likewise, a real dev, if I said, hey, make a landing page, they'd be like, why, what, how? I don't care. What do you mean I have to talk about them? I made a product. I want to talk about me. Well, that's not how you write landing pages. Now, will they make a result as good as I was if I was running it? No. But suddenly it gets to where you're 80% overlap on skills just by asking. That's a superpower, man.

Yeah, for sure. And, uh, and one of the other things is like, on how do you spot whether or not an agent or an AI is doing the right job is if you are, let's say you're a marketer, right? So if AI were to make a mistake or if it's not optimizing for the best output, you can really, really quickly spot it and then you train it and it becomes better and better. And I think at the same time you have this great opportunity to learn very rudimentary coding skills and developer skills, right? And in a way, like you said, like you're kind of meeting halfway.

Exactly. And that ability to level yourself up is all the time. You've got a founder who goes, man, there's me and there's the CTO. Who does sales, man? I don't know how to do sales. Or who does customer support? Or who does whatever? Well, you know, with an agent to help you, You get 80% of the way there. It is way better than you trying to do it yourself. All right, we have a few minutes left. Uh, I do want to respect everyone's time, but any last questions? I will once again— I'm going to go back to the, to this, to here, to this. The next session, which is, I think it's next Tuesday, isn't it? It's next Tuesday, same time, different link, I'm sure. Same time next Tuesday, we will be talking about, I guess, speaking of like some specifics, how do you evaluate quality of outputs? How do you make your moat you? How do you become a, uh, founder whose story Motivates people to buy and how do you distribute it? Most importantly, because I gotta tell you, when you check out CrowdTamer's YouTube channel, which is forward slash at launch today. Did I spell it right? Possibly. Hey, you did. Uh, how many subs do I have? 188. Uh, YouTube made me about $200K last year, right? Not because of my subs. My organic reach is not the thing that makes me money. It is ensuring distribution to the right people. So it makes me money. And if you're like, hey man, I would like a YouTube channel with 188 subs to make me a couple hundred K. Come next week. I'll see you then. As for Buddy, Buddy can go pound sand, man. Buddy, that localhost, a localhost you did me? Fine, localhost can go screw itself. This is supposed to be in the cloud. Fine, there we go. There's the landing page, right? Which is this great design. Does it read like AI? No. So at least we'll accept that. I'd also note you didn't even— oh, this emoji have a mustache? Oh, I guess it does. Fine. Also, that alt text is terrifying. I am watching you. Are you, buddy? I'm scared now, man. All right. Hosted a local host. I'll take that as my booby prize for the day. I'll see you guys next week where I'll discuss the content engine that we have built What have we done? We built $270-something million marketing engines in the last 7 years, and all of them are backed by these core ideas. I'll see you guys then.

Awesome, folks. I hope you had a great time. Uh, I've shared a link to Crowdtamers in the chat. Uh, do remember, we'll be sending over a recording as well as all the skills that Trevor talked about. Do not miss out the next, uh, workshop in the series. I hope you have an awesome day, and if you do feel like networking, you know, meeting founders from across the world, do stay back for a little while. We do have our networking lounge right after this. Have a great one.

Take care, y'all.