Founder's Guide to AI: Converting Data into Business Value with AI

Hey everyone, I hope you are having an awesome day. My name is Daniel, I'm a part of the Founder Institute marketing team, and we are back here today with what is the third and final workshop in our three-part workshop series, AI for Startup Founders. As always, I am super, super excited to be joined here with Trevor. Trevor is the founder of Crowdtamers. He is a great friend of FI. He's a white coder. He's launched startups, he's helped startups grow. So there's a ton that you can learn. And today specifically, he'll be talking about data and how AI can help you extract incredible insights. Before I let Trevor take the stage here, just a couple of quick things to remember. If you have any questions for Trevor, please do drop them in the chat. I'll be back on stage in a little while so that we can answer all of those questions. I do see that, like always, we have a super international audience. We have people from Chicago, Toronto. I see people from India here, Portugal, Berlin. Awesome, awesome. So Trevor, we have folks joining from everywhere, all here to listen to you. And finally, folks, if you are enjoying what you hear, what you see, there's an emoji button at the bottom of your screen. Please do let us know how you feel. Click it, you know, clap, thumbs up, whatever it is that you're feeling today. Trevor, the stage is yours.

Awesome. Thanks as always, Daniel. So appreciate everybody making time on a busy Tuesday to come talk a bit about, I think, the least exciting part of AI for Founders, right? AI for Founders about the data. And one thing that I know is the sexy part of AI is the agents and the automation and building cool things. And we've talked about those, right? We went into, uh, at the beginning, we kind of talked about you are the moat and trying to make your features what makes you unique. Isn't what's going to move the needle anymore. You've got to get AI native. We talked about the sexy bits already. And then we talked about the specifics of, like, CrowdTamer's how we build million-dollar marketing engines. I was just thinking while Daniel was talking about how I've been working with FI for a while. I think it's 3 FI startups. I've helped them go from C to B or somewhere around there. and, you know, do $15 or more million in revenue. So, uh, the insights and the, I guess what I'm telling you here is both true in the specific, but also true in the general and based off of what I've worked with and built for FI startups and other startups over the last decade. Although these slides, I think I finished making them, uh, yesterday. So they're about as current as I can. In what's available and what's capable for you. And of course, I am presenting this as I always have been for this sequence through impressflow.ai, which is the tool that I wrote so I can make presentations happen faster because I don't want to spend a week or two designing slides when I could just have them be live instantly. The slide design might be a little bit rudimentary, but the ideas, I hope, are still solid. So today we're talking about the least sexy part, but in some ways the most important part. Which is when you have agentic AI, when you are doing marketing in the manner that CrowdTamers has done and been successful about dozens of times in the past, you need to have data to steer. And you have so much data these days, it's going to overwhelm you. There's an infinite amount. There is no limit to how much data you could get. And it's really easy to tie up. One people, several people, your whole company into understanding what's in all of those tabs that you know you should be checking, right? Just the core basics of what I want to check every day. My analytics, my Search Console, my YouTube Studio data, my LinkedIn ad data, my YouTube ad data, my HeyReach outbound data, how my email newsletter performed, and finally what my CRM says I'm going to make this week for revenue. And I'm running a relatively small business. When you have 10 salespeople and hundreds of deals flying around every which way and thousands of visits to your website daily, it gets insane. And there's a reason why they employ data analysts at large companies to do this, but you don't need a data analyst because you with Claude's help, can be the analyst. All you have to know is what it is you need to know. Because with all this data, the tempting thing is to say, just show me charts, let me see lines, let me— there's metrics, right? A lot of the data you can pull in is basically vanity. And one of the key things we have learned at Crowdtamers over the last year is AI can do an astounding amount of the analysis and insight information if you use it correctly and if you don't trust it. Because AI is a brilliant, amoral, hyperfocused intern. And because it is brilliant and amoral and also hyperfocused, it'll make up stuff when it finds it more convenient to make it up than to go find the truth out for you. So part of what I want to talk to you about today and what I want to give you the skills for today is how to make it where your AI is dependable and believable in the outputs. Now, if you hang around till the end of this, or if you're watching the recording, skip to the end, I'll share with you the links to the skills that CrowdTamer uses to drive our analysis for all of our clients. I would note we do this, and I recommend you do too, AI takes a first pass at the analysis. You still have to look because AI will make stuff up. AI will draw conclusions that aren't actually supported by data unless a human shows up and goes, hey, why did you say that? And it goes, oh yeah, you're right. I made the whole thing up. So what matters now, and one of the big changes in how we use software since really November, is it is much faster these days when you want to understand complex data to not use the tools that have your data at all. There are 3 big shifts your mind has to make to embrace AI agentic work in 2026, right? The first one is, we talked about it the first session, all AI work is iterative now. The first time you ask AI to do something, it will suck at it. It'll get better every time thereafter if you steer it. The second thing to think about is, uh, we are, all of us, managers of context window now, right? The more I can control what my AI needs to know for this task, the better a result I will get for it. And the third thing you have to understand is that AI is making software transparent. I've got 2 or 3 tools right now that my primary use of them is no longer at all in the tool. Notion, for my company, I haven't logged into Notion in the Notion app in, I don't know, a week or more. I transact with Notion through the APIs. The same with a lot of my emails these days. I transact with emails through the AI, through the API. That transparency means you have to be comfortable pulling the data where you need it so that you can now use it in ways that maybe people didn't initially think about when they were building these tools out. So how do you know what to ask? How do you understand what combination of things ends up well, like peanut butter and chocolate, and what combination ends up poorly, like toothpaste and orange juice? It's a matter of having spent some time finding out what matters for you. So let's talk about data sources, what they tell, how you want to combine them. Right? When I say Google Analytics here, I mean any analytics platform. I'm using currently right now Matomo, Google Analytics, and Umami Analytics all kind of interchangeably. I find that the new, it's now like 18 months old, Google Analytics still remains extremely hard to use for a lot of use cases. And, uh, for particularly client-facing stuff, I may be using other software than just Google Analytics, but, you know, make the replacement in your head, right? Before anybody does anything on your website, you have to understand that the session length, the pages they look at, what people are doing on your website is the clearest indicator you have of whether or not they are valuable people for you. You might get 100,000 people coming to your website. If their session time is a quarter of a second and they leave, they don't matter. So Google Analytics tells you where they went, how long they stayed, where they went next. That is kind of a foundation of, is your website as a brochure, as a sales tool, doing its job? But a lot controls what makes the website work. Because how I'm bringing people in is going to directly influence how these pages perform. So you want to start to ask some questions around, on a given page, what's driving growth? Is that traffic good traffic or bad traffic? And then as I am spending money versus spending time, to bring people to my website? What's performing better for me? I just turned on the new ad campaign for CrowdTamers, uh, I think this last week, maybe Wednesday, pretty recently. And one of the things about it was that it's the beginning of a new ad campaign. And because it's the beginning of a new ad campaign, it's not like the traffic isn't super well tuned yet. So when I do something like I say, Uh, here, let me actually do this. When I say, hey, we're doing a good job with overall traffic numbers to CrowdTamers.com, perhaps. But then when I look at in Google Analytics, something like my, uh, traffic acquisition, and I can look at this and say, hey, my paid video, right? 27 visits. That's the new campaign we turned on. But average time per session is 0 seconds. Like, this is clearly traffic that isn't very good yet. So I want to work on how do I make that traffic better before I worry about upping my spend, before I worry about improving my landing page, before I worry about anything else. I want to draw insight just from here. So you want to ask these questions, but then You combine your data together, right? In Google Search Console, I can see things like what's being talked about right now. So going in there and seeing what is effective for, for example, again, I'll show you CrowdTamers. And I can see, am I ranking? Am I showing up? And if so, what am I showing up for? What are the keywords that I'm popping for. And I recently wrote some articles about OpenClaw and Obsidian, so I'm starting to be found there. But most folks who find me are searching for my brand term. I'm not surprised that I rank for that, right? But then we begin to combine this with this, and previously I'd have to open up two spreadsheets, hold it in my head, and understand. And then I might go, hey, beyond this, I want to understand What trends are growing? What am I being found for that no one's clicking on? What am I being clicked on that's higher or lower volume? Are pages winning or losing in search? Then we say, we can intersect these two things together and say, this has a lot of demand. When they come to my website, they perform poorly. They don't spend a long time here when they've gone to this URL. So, Ranking or not ranking for that maybe matters. Maybe it doesn't. Maybe I need to make the page better. Your judgment has to come in here. Now the AI can help steer some of these things, right? So then we go into YouTube Studio and that's where we can see what of my content, of course, that I published on YouTube. If you're following my most recent, uh, the previous session, right, on how you do YouTube content engine to build and grow your company. And then we see things like, okay, I've got data here, like 25K views, big spikes. I just turned this campaign on on this channel. And, uh, I can dive in and say, which videos are getting views? I can dive in and manually say, how long are they watching? And the answer is Where's my average? There you go, 16 seconds. That's 27 seconds. Like, okay, these are getting some views. People are paying attention. I can download this CSV and export it or even connect to the API. But then we begin to look at a three-way filter, right? Because YouTube drives search demand. People who are searching crowd tamers, a fair number of them are coming from this because crowd tamers as a search term is taking off recently. Right. I turned these ads on because look at that, CrowdTamer searches spike right around when my video views are spiking. There's a definite connection here. So which videos are starting to drive search as I turn videos off and on? Manually keeping track of it in my head is going to be an unbelievable mess to try and do, particularly looking back 2 weeks, 6 weeks, 12 weeks in the past, but it all sits here and the AI can jump in and analyze now a 1, 2, 3-way different view where you turn this video on, which drove demand for CrowdTamers, which then brought people to this page, and then they were likely to book. This is what you get when you start to combine these sources together, is a way fuller picture than just, oh, my YouTube is not doing great this month, I should do something, or, hmm, search is up, I wonder why. And you can begin to say, are people rewatching my content? Now, I've mentioned in the previous workshop last week, when you have good content, when you have content that people tend to rewatch or tend to watch more than half of, it drives organic discovery. So which videos are driving that discovery? And the organic discovery, how much are they watching? Right? Are there particular subject lines or topics that are doing better? All this data begins to inform bigger questions, right? I've moved now beyond just, hey, what video works? Because YouTube will tell me that. And in a big company, you may have somebody who's purely optimizing video. But that's actually a waste of a lot of effort because if I got my video numbers to look great, but they don't drive my revenue at the end, then it's worthless that I've optimized my video channels. I need to look at this holistically so I can start to say video numbers are up and they drove this increase in video or in search discovery, which then drove this increase in landing page performance. Now I put these three things together. I begin to say, who was seeing these videos? Where should my marketing steer to? What should my product be developed for? How do I position my entire business? Right? When you begin to analyze all these pieces together, something as simple as which YouTube Short got the most view-through rate becomes, oh, I need to spin up new landing pages and new positioning for my entire business because this is an easier path for money. That's only when you combine 3 pieces, right? Then you start to add LinkedIn and HeyReach outreach. I say HeyReach here in particular. Because I like their API. Phantom Buster has it. Alfred has it. Lemlist has it. You can find a lot of ways to get data from an outbound tool. Any of them, as long as you can pull down all of the raw data on all of your new raw followers, will work. I just happened to find HeyReacher's API is really convenient compared to other tools. I don't have to bother to download stuff on LinkedIn or from the tool directly. I can have my AI interact with the API and solve it for me. So as an example, and I'll show you some anonymized data of this in a moment, but in HeyReach, being able to say something like, hey, 1,300 connections, 372 accepted, right, a 30-ish percent accept rate. Okay, that is what you'll see in something like I'll log in real quick and show this to you. I told Daniel at the start of the presentation, no, I won't be sharing any other tabs than this one. And then instantly I've ignored my own claim and I said, nah, let me share every tab instead. Okay, here's stats from my current campaign that's running right now, right? My connection, my connection except pretty bad right now because I turned off the LinkedIn ads because I'm proving a point. If you advertise your content before people see outreach from you, which I think I was doing back here. Yeah, there you go, right? 40%, an extra 60-odd percent in accept rate. So over 2 weeks, 357 connection requests led to 62 replies and about 20 meetings booked. Whereas if I do here after the ads were off, right? I mean, tailing down, but same number, 55. And if I go and add in this week, which we're early in, but I go and add in those 2 weeks here, right? 30%. So dropping, right? But anyway, this gives you the high-level view. This tells you that you sent connections, that they were accepted. It may tell you what campaign is working, but it's hiding a bunch of data, right? The data it's hiding is things like Of these replies, how many were positive? Of these positive replies, who were they from? And even what job titles, what geographic patterns work best here? You can use your LinkedIn outreach campaigns to understand things like what countries are most likely to buy my product? What type of messaging works? What type of positioning around the thing I built is functional? And then keeping in mind that outbound campaigns also drive search visibility, right? I get an outbound message from somebody. I might google crowdtamers.com and then find CrowdTamers. And from there, sign up and book a call. And you can track if you isolate channels, right? Oh, this drives this much new organic discovery. So as I'm adding in my outbound campaign here, it's going to give me wins in organic marketing. But which one is most likely to drive people's interest to go check out my website? All these interact together. And when you get all this, you might find something like this, right? Short characters, short messages, we can prove, have this much better reply rate, but especially in Germany. To jobs of this title in the US, to jobs of this title, a different campaign will work better. That's really hard to find out by looking at the raw data. So all the dashboards tell you is the high-level view. They don't tell you how things fit together, how they intersect, and how you can understand a lot more about your data with some help from AI. So any of these platforms you run, right, I'm actually going to go, I'm kind of belaboring the point now, right? But any of these platforms that you run, these specific searches will give you platform-specific knowledge. And I do recommend you do that because you want to pull together your isolated data. But then you also need some kind of meta-analysis that looks at everything and says, based off of all of this stuff, Now we're going to draw a bigger conclusion, which is, we're going to talk about now, right? When you bring all these pieces together into meta-analysis, you get from content into keywords into understanding what you actually sell. So this, uh, this is a very it-focused a Fractional CISO brand client of mine, right? We looked in what Google Search Console shows for visibility. We looked at what's published on his website. We look at what's published in his content and we see, wow, there's a big gap there, right? No one's talking about this. We can say, hey, we've got a video that people love to watch and rewatch. It's high engagement. That's a pretty obvious choice. To make an ad. And then when you put those together and you run that ad, you're going to see, hey, the content you've been making around, hey, don't be scammed by people, that's not what's effective for you. What is effective for you is creating a, uh, unique positioning to your it directors that says, hey, we're going to find from there a, a new approach around who you want to market to. And that comes from pulling together YouTube data, ad data, website data, see what's driving folks to sign up and actually be part of the website. All the way down to the point where now, for specifically him, we know this is the way you want to construct a video headline and you want to construct your content, right? Name a job title, talk about something that went wrong, talk about why it's bad. And if that's your headline, you will get more views. And not just more views organically, but people who see the video who are targeted, right? If I say it directors, uh, who don't control their company's use of ChatGPT are going to go broke. There's a headline, right? Then we target that to the IT director role. And then we know they're going to be high engagement. When they're high engagement, they're more likely to book a call. More likely to book a call, you'll make more money, right? Any one channel, it's really hard to pull this data together. You bring it all together, you get the picture you need to understand everything. Let me talk about a campaign I'm running for me. In Heyreach, right? From there, we can look and see we had 126 conversations in North America, in the US. But when I started to look at it, I went, hey, like a third of these are negative, or maybe even almost 40% are negative, which is pretty high for me. Why are they negative? And what I found out, this is outbound for CrowdTamer, the content agency, right? Me, right? Do you want to come? Uh, this was a— I was offering some kind of content marketing audit at the time. And hey, would you like an audit? And the answer for a couple of the mattress stores was, heck no, no interest, you're a scammer, buzz off. Okay. But in some areas they did better. And why is that the case? I was unsure about. And what I looked at was There was, there were two things happening here that I wasn't aware of. One is my LinkedIn ad spend was not evenly distributed geographically, because of course, when you tell LinkedIn, hey, I want you to optimize for cost per view, it's going to optimize to show your content in the cheaper parts of the US where the views are less expensive. But if you're then doing outreach in LA or New York or some places where the There are a lot more eyeballs, but they're also higher value eyeballs. They have to pay more for them. Then I hadn't actually warmed people up. They didn't know who I was before they got my content, which is not the whole point of the content engine. So I had to reallocate how we were budgeting our ad spend to show it to the right people. And then what I was looking at is I've got some people we were driving this content to, they were checking us out on our website using a tool like RB2B. You can see who's coming to your website and then they get outreach from me later on. Those folks were the most likely to reply, right? They saw my content, they were interested enough to Google me, but they maybe didn't book a call. No worries. I've got you now. Inbound and outbound response, both of them put you at the front of the line to get outreach from me right away. If you're coming into me and you're not driven by outbound, I should be doing some marketing to you elsewhere and remarketing covers that. Uh, and then if I'm sending out and not getting replies, well, obviously you stop trying those areas because they're not worth the effort. So, okay. Um, another example of Search Console plus YouTube or Search Console plus Website Analytics or any of these things. If you've looked at me on LinkedIn recently, you'll see I've posted a couple of times about Open Claw, which is the agentic AI harness I talked about in the first workshop, which continues to be super relevant to people right now. And my two most recent blog posts are actually my three most recent blog posts, all about this, right? This one needs a better graphic image, but I'll fix that today. That's because in Google Search Console, whenever I post about this, I see a big uptake. Now, the thing I don't know yet, because this is only what I think 7 business days worth of content, is can I convert these people into actual leads, clients, interest? I don't know yet, but this is an example of Search Console data plus Google Analytics data. Because I need to see, right, what's your time on site when you come here? After you read this, do you subscribe? Do I get you with the pop-up? Like, what makes you want to give me your info and persist? It may be that I can get traffic here very cost-effectively, but if none of them want to hire a marketer, that's not worth my effort to keep writing this stuff. Or to use the YouTube content calendar example. You can look and see what's the rising queries in the video tab in, uh, Search Console. Because I can see here, right, like I've obviously got big spike on my brand name, but I can in general see what's trending up, what terms are showing up more often that are leading to my site. Or I can do a general investigation using Google's Keyword Tool to see not just what's coming to my site, but what's in general more interesting. Do I want to make content on that? Right. And then do I have video already for these queries? Because if so, run an ad, get the views. If I don't have video, let me prepare it. One of my clients yesterday, yesterday, they're a no-code automation agency. They build like integrations for big companies. And they realize that there's a ton of demand in a couple industries that the AI revolution has kind of left behind because of HIPAA compliance and legal requirements and these sorts of things that they could serve. There's growing search traffic by these industries trying to find answers. No one's made videos around this. So of course we recorded two video sales letters yesterday, and we'll do two more, I think, on Friday to cover these industries and say, hey, This is what we need to talk about. You've got a 5-minute video, that's the sales letter, and then a couple of 1-minute videos that are shorts to drive interest. Then what we have to see is, wait a month, what do they do on site? Do they convert? When they show up in your CRM, they came from your site, they came from YouTube, cool. Do they end up giving you money? Put all the data together, you're gonna understand what you actually need to know to make business decisions. That's not formatted correctly. Okay, so my live demo ended up kind of being spread out across this, but let me describe how you do this. So a couple pieces we want to pull here from— I'll do CrowdTamer in particular because no one's going to fuss at me for sharing their data. If I go look at my visible performance, right? Here I can grab, export the CSV. I'm going to download. This is my 30-day— this is my 3-month data. I don't need to explicitly grab my 28 days data here. There are some tools where you'll want to grab them separately because you're going to get different info. This just shows performance across my Content. The other thing is I do want to see indexing in general, how I'm doing. A lot of my pages aren't indexed. That's because a whole lot of my pages are just single one-page files. Sorry, one-file pages, because WordPress just posts every image as its own separate page. They don't need to be indexed. So this is fine. If I wanted to, I could export this also and see how it maps up. In my case, Do it, you know, once a quarter or so. You don't have to do it that regularly. Then that's kind of showing my, uh, my demand, right? What's visible out there on Google search and the spiking visibility of my brand that is led in part by what's going on here. So I've got my daily chart for the last 28 days. I will download the CSV of that. And then I can also take the last 90 days, which is pretty flat once I turned off the ads from this channel, and have a separate view for that because you may want to have the AI look at both these as separate contexts. And then I'm running in my website analytics. We're going to do traffic acquisition. Hit the share button, which is a weird place to go find it, download the CSV. And then I don't think I have lead acquisition currently hooked up properly here. So that's not reported back into Google Analytics. I have it captured in Umami instead, which is a different analytics tool. I do also have my conversions here. E-commerce purchases are not how I fund my my business, uh, you can see I made a whopping $168— $158, sorry— this month. All right, if this was something that was important to me, seeing what sold, what people bought it, all this information might well be worth exporting to analyze separately. In my case, it's mostly a curious thing because I don't really care about the revenue I get from this. Indeed, I think aside from Validate First eBook, which I think is still priced at $8. Everything else here is free or pay your own price. So that I made $156, $158 baffles me because I've only sold 3 copies of the thing that actually gets paid. People were just donating money for some of these other things, which was nice of them. Thank you, I appreciate that, but it is not what keeps the lights on. So, uh, and then the last thing I can run here let me do make sure I'm logged into my ad account and not somebody else's. Is YouTube ads run in Google Ads? Because why not? And as a result, I've got this new campaign I just started, like, with a fairly modest little spend. And, uh, I can do— we'll do CPM. There we go. I can look at Total views. There we go. So this campaign here, I can go look at the campaigns with my little $15 a day spend on this campaign, which is the one I turned on live last week in the live demo with us. And, uh, have been spending not the full value yet because it's still warming up, but we've been spending— what's our total spend here? $7 on this campaign. And then the other one that's been around a little bit longer and is fully warmed up is this one, which we're spending closer to it. So what I'm going to do is in the ad groups, going to download the CSV. Nope, that's the wrong one. When you download this CSV, that just downloaded the chart and grab the ad level details as well for this campaign. And then go back to the FI campaign, do the same thing, download ad-level and creative-level CSVs. Now I've got YouTube Studio for 60 and 90 days. I've got this ad campaign, which only turned on on the 26th. And then I have— allow— then I have the other campaign I was running. Let me this one, which has been on for a little bit longer, so I should adjust this window, but also not super long, and download that and this. There we go. Now I've got my ad-level data, which was driving visibility at the beginning of the journey, along with my search performance data, which is the other start to the buyer's journey. Then I have my analytics data, which is what people did on my website. And I've also got my YouTube Studio data, which is in-depth information about what people were doing with the content that my ads were serving. All this pulled together, I can now say I've got a giant mess of files in my downloads folder, which All right, this is full screen. Uh, what I'm going to do is— now you can do this a number of ways. I will start this with, uh, running my— we'll do this real quick— with running, uh, in Claude code, or sorry, Claude in the desktop. Uh, you can do it with Claude code. You can even do it in the browser if you want to. Uh, oh, I have to switch this way. There you go. Okay, uh, you can run it on any of the platforms that you want, it's fine. But you're going to grab all of your files, which we can see here. Normally I do dump these all into a folder to organize them a bit, but you can see I've got a ton of data from today scattered around. Any case, this is the presentation here. So I can start with, I have some skills declared and those skills, let me share with you in a moment here, are ones that you will get for download momentarily. And they are marketing-specific skills. And they include things like YouTube Shorts analysis. I want to grab— now, when you have this properly installed as a skill in Claude, you don't have to do this, but I am going to grab the path for the purposes of having this ready. I have a YouTube Shorts analysis skill here I'd like you to use to tell me what of my content is performing and if there are any topic clusters I should make more of. And then I have that skill there to steer it. So there's a ton of data here. It's going to start to churn through and use that as the basis for, uh, what it's going to tell me. And in the marketing skills pack that I'm including for download, You've got these skills and a number of other ones too, right? Uh, I see we've got YouTube Shorts analysis. We have a couple other ones. I have to find them in my root. I don't remember where they are in Obsidian, uh, but I have them here, right? Data presentation, marketing data strategist. All these skills are here to walk you through individual like tell me more about this. And then at the high level, the data strategist is what you can run at the end to roll them all together. Now I'm going through this one step at a time, which is not a— uh, you want in-browser? Interesting. Give me a second, it's being shirty. Uh, use desktop commander. I'm running this in the desktop app. There we are. Okay. The smart way to do this when you're going at more scale is to daisy chain these together and to have Claude write a script that says, hey, I'm going to first do the YouTube analysis and then I'm going to do the email analysis, whatever order makes sense for you. I'm going to run all of these, save these all to a file somewhere, and then I'm going to run the meta-analysis on all the analyses to give you— yes, allow— to give you the information that you need to draw conclusions from all the individual pieces. Now, if you don't know how to write scripting, and I don't know how to write scripting, you can ask the AI to write the script to teach you how to for you, right? I'm showing you the simple version now because I didn't want to show you a fully built-up, uh, set of, uh, workflows here because I'm trying to show you how to begin from the beginning. So this is going to run and take a few moments to generate this. I will leave it be while it does, and, whoop, and, uh, I'll come back. So what we're doing here is pulling together, in this case, all the raw files direct from my workspaces, right? Certainly when you narrow the focus of the AI, you will get better quality outputs. But in general, for Claude, you don't need to because the context window is big enough. And they are rolling out in the near future. In the near future, they'll have a 1 million token context window, at which point you'll never use it all. So once you've got that, right, the prompt is quite short, right? I wrote— how many sentences was that? This many. But then also, if you think about it, this is the prompt, right? This tells it what to do. And when you give it that as the repeatable chunk of the prompt every time, it's easy to make durable lessons. I mentioned how the 3 things you need to learn to become truly AI native. The first one is all AI is iterative now. When you download from the link I'll be providing this skills pack and you go, Trevor, that sucks. I want to know some other thing about it. Edit it, or better yet, tell the AI to edit it for you. And then the next time it runs, it'll get results more like what you want. But if you don't make a skill out of it and edit the skill, the next time you run next week or next month's prompting, you won't have the record of how you want it to be. So think of this as the executable file for your AI to run. You want to iterate and improve on to make it perfect for you. The stuff I'm using and the stuff I've shared in the skills file download is an 80 or 90% effort that you're going to want to tweak further from there. Okay, I see Claude is, uh, taking its sweet time. They must have— they had an outage yesterday, and I guess in recovering from it They're still running quite slowly, right? So you write the prompt, and the prompt mostly includes the context you share and the skill file that you load. I would note, absent deeper info, this is no good, right? But you can make this work if your skill file is sufficiently in-depth where you say, when I mean top 10 pages by traffic, I mean, top 10 pages that have a conversion event or that do this or have more than 12 seconds dwell time on site or whatever it is. And then you can do lazy prompting and the AI will take care of it from there. Or you can be more specific upfront. I tend to find most of my work is very repeatable. So I build good skills so I can do lazy prompting. Then we're going to get a result, which I'd hope would be faster than this. But as is always the case— hey, there we go. As is always the case, um, oh, I pulled some of the wrong data in. Okay, um, some of this is using somebody else's data because I was not looking properly when it pulled together. Um, On the awareness campaign, that's my campaign. Uh, this is telling me, hey, that much I can see is already pulling out some information. Hey, it's recommending maybe your cold open or your hook sucks, right? People who don't skip this watch the whole video. But the video starts badly. Record a new intro. Maybe it'll do better. Easy, right? And then, hey, it's roasting me on my own titles here. Okay. So it's also saying something I already called out, right? Session time from this is no good. And it's giving me recommendations for what to do. In my video content creation. So that right there, by pulling together not just what's in YouTube, but also it has what's on my website, it's able to say, hey, your website clearly either isn't registering properly, although it is, or the people you're bringing are not performative. Okay, that's its own problem. And there are two possible reasons why. And I know, because I know my website, Right? It might be just the website loads too slow. It might be some other reason. Testing and finding out why for this ad campaign is its own separate thing to solve. But something to pay attention to is I could say, based on that info, use this skill. Let me grab— I know where it is here at least. I'm going to do this and we'll copy. In my— this is my Downloads folder. You can't access my Downloads folder. Hang on, I've got to move it. There's a sandbox set up here so that I can't have this. Let me copy and paste it here. There we go. In my Documents folder. To tell me what types of traffic aren't performing on my website and maybe what I should fix on my landing pages. Right. It's got a bunch of data. I can ask it now to look and see, and it'll probably take a few moments to go read it. So while it's doing that, let's finish up talking about this, right? Because it's given me some data. And I would note in this case, I gave it some bad data because I include somebody else's channel. But you need to bump into it and say, hey, you said this, is it true? Now, when I'm doing client reporting, one of the things that we do is we run the report twice, identical report twice. And if the numbers in the report aren't the same, stop. I'm old enough to remember, maybe some of you guys are, the movie Minority Report. Minority Report had 3 different psychics that were able to predict the future. And if there was a differential, if 2 of them saw the future, they accepted it was true, even if all 3 of them didn't agree, because the odds that all 3 would hallucinate the same thing is pretty rare. It's the same with AI. If you get the same numbers twice, the odds are vanishingly tiny it's a hallucination. So if you run it twice and you get an error, run it a third time. And you should have either all different numbers, in which case stop, you're hallucinating like mad, or you'll have two that agree. Awesome. That's what you can begin from here. It's one of the easy ways to offload the "is this real" out of your head and onto your tooling. Having it run that way means it will almost certainly be true data as it talks to you. So let me just see. Still thinking. Okay. Um, then the catch here is what I just showed you, you can do by hand, takes a few minutes, right? But on the other hand, you can also get a coffee, have a chat around the water cooler, whatever, while it's thinking, and you'll get the results in a few minutes. But what's even better than having to manually do it, at least in my case, is have something set up that will do it for you. And there's two ways that I know of that you can easily set this up to run repeatedly. One is Claude just rolled out scheduled tasks in Cowork. Now it's currently in the midst of thinking and attempting to move it over here will probably kill it, but I can say, I want you to do this skill, and you could set all the dependencies you want for how the skill works up. Every Monday at 6:00 AM. So when I come into the office, I'm going to have my reports ready from last week. And then as long as you left your computer on and left Cloud Cowork signed in, in the morning at 6:00 AM, it'll go do it. You can also use command line tools. You can either write them in— I clicked the wrong button. Either write them in Claude code here, you can write them in Claude code in the terminal. And I know I have an absolutely psychotic terminal color. I've had real developers make fun of me for it just this week. Why is your terminal pink? It's because that way I know it's a Claude code terminal and not an OpenAI terminal or a Gemini terminal. But as it may be, right, the Results coming back here finally on my website, right? So paid video doesn't work at all. That is correct. I already talked about that. The other one is organic search hangs out and doesn't do anything. Okay. I mentioned I'm driving up visibility with terms like Open Claw and Obsidian, and I'm being more discoverable. And of course, CrowdTamer, the search term, has also peaked, but conversion rate's not great. So organic, social doing great. If you find me online through any of my content, that works well. That's why I run a content engine. Okay. So this information here gives you immediately information across all those different channels I pulled the data from and starts to make recommendations for you as well based off of what the skill says, right? And using that as the basis for a regular reporting habit can make you really start to see how all your channels combine together effectively. And then you start to daisy chain other pieces in. If you were in the first workshop, or if you get the recording for the first workshop, after watching this one, among the skills that I gave away in that one were skills specifically around thinking, guided thinking, how to activate, do things like give you interesting ways to solve problems. There's a, a specific one in here somewhere that is called Challenge. And the idea is it actually pushes back on you and says, hey, why aren't you doing this? Right? These skills coming your way means you can ask the AI, hey, what am I not paying attention to right now? Challenge me. And it'll go through the thinking and walk you through it. So you start to say, great, I've got all this data. Who knows where to find it? It's in YouTube. Go to F— making analysis to then have AI say to it, hey, this is what I think I need to do. Like, I can say pretty easily here, I know my homepage load time is really bad right now. It's something like 15 seconds to first paint. Should I probably fix that as part of addressing why my content and my website performs so poorly for bringing in new leads? And then the AI might say yes, it might say no, but I can have it challenge me and say, yeah, that's probably it. Ah, well, okay. Yes, maybe no. Right. I can now have it as the thinking partner to help me make better decisions. So make it regular. Think about the patterns that are going to work. Right. I already mentioned, and I showed you one right here, right? I've got the, the time on site is so slow for organic video compared to the time it takes to load my website. Might that be a problem? Might that be something to fix? AI goes, probably not. Okay. I can read why. One of you are not watching me. Right. Have it challenge you. What's not going right? Having it as a strategist is a, a, a mixed blessing because what you tend to get is fairly mid-strategy. You won't find truly wild creative stuff from AI unless you push back on it. I had AI for a client at the beginning of this year recommend to me, I went, hey, I want to do an event for this company in, in New York. For a bunch of European guests. What's something to do? And it gave me some very bland ideas, and I went, no, those all suck. Come up with something more interesting. And we bounced back and forth one or two times, and it came up with the idea of renting out a New York bodega. Pay a bodega owner a couple of grand, rent it out, have an event there, serve, you know, the iconic New York street foods, have a cool event there. Right? Don't even need to serve alcohol if you don't want to. They don't have a liquor license, so you can't. These sorts of things are very non-traditional, but I didn't propose that first, even when I said make up something fun. It was like, why don't you do a, you know, go up to Top Shots Golf? I'm like, that's not different. That's pretty easy to do. So push back on it and you might get better strategy out, but your human taste test has to show up. You have to come up with an idea. That's different with the AI. Don't expect it to get it perfect on its own. And then you want to look for your counterfactual, the things that are hidden in the data that you might not see otherwise. You can push and prompt at the skill level, at the strategist level, all throughout it. Hey, what's something that you didn't reveal? What's something that wasn't obvious? But especially on anomaly hunting, validate it is true before you believe it when it says it. Because if there are no anomalies and you say, what's the anomalous thing, it's going to lie to you and come up with an anomaly because it wants to make you happy. And then the whole basis of what we've been talking about here is how do we combine these pieces together into a whole, but you may have email marketing or you may have, you know, sponsored UGC or whatever as a channel you want to look at separately and then compare across all the different pieces. So I've already talked about this, right? AI will lie to you about this. The best way I know of to make this go away is to ask it in a separate chat, do the same thing and see if you get the same result. You can also push back in that chat and say, great, I see your report. Cite me the data for every conclusion you have and name me the spreadsheet and the row you found it in so I can go look it up myself. And then I might be like, oh, I can't find all of those for you. Ah, yeah, I kind of made some of those up. Okay, cool. Now we push back. We found that out, right? Also, the more info you can put in on how each individual pieces fit together, the better the overall strategist view will be, right? If your site was down for a week and I was like, you didn't make any money last week, yeah, that's why, right? Giving that information to it is crucial for it to understand. And then your choices for what you import and what you export and the data you share and all of this will steer it certain ways. If you run ads on Pinterest and don't report on them, with a tool like this, then it's going to tell you that your successes all came from somewhere else because it only reports on what it knows. So do the best you can, right? The 3 things you have to know to be AI native: AI is iterative. Your job is to manage context windows with AI, just like you manage new onboarding for an employee. AI is always new. Every time you open a new chat, it's a new employee. Manage the context window like you would onboard a new employee. And then, right, AI is iterative. You have to manage context windows. And AI is transparent. It makes all your software go away because it becomes one point for the center of it. All this reporting is pulling together for you. Then do make sure that both in your case and in the AI's case, you are not only pulling a month's data, right? Pull a month, pull 90 days, pull a year. Run those separately, and then start to see what knowledge comes across all those gaps. Now, I didn't do a great job of it for this presentation because I have a client's name in the files, but I didn't share any of his PII data here. And I would note, when I'm doing things that are sensitive— CRM data, right, customer contact data, any of those things— I am not pushing these to Claude. I run these locally on my machine. I've got a Mac Mini with 64 gigabytes of RAM that can run a bunch of models very nicely locally and give me the report. And then that report I can use for the meta-skill analysis to say, great, all the PII has been removed. This is a generalized report now. And all the data sat only on my computer. So my customers, I don't know, phone numbers of their customers aren't leaking up to the cloud. This is more controlled. If you think of AI always as the new hire until you've really built the skills and the SOPs that if you gave a new hire on day one, they would never make a mistake, right? You have to hold their hand and bring them along, but it'll perform extremely extremely well. So this is the end of 3 hours more, 4 and a half hours we've been spending together across all these sessions. There's a couple key things that I want to hit on and encourage you, if you haven't watched video 1 and video 2, or even if you have, to refresh yourself. AI is a partner. If you let it be one. My younger brother who works at a big company this morning was like, Trevor, you've been talking about AI a bunch when we do our family calls every Sunday. And you were like, I've just started to play with it. You seem like maybe you know a little bit. Tell me about what I should be doing. And he was at the beginning level of just like, it's a chatbot. I'm using it as a chatbot. Right? It could be all the way up to, I need you to write this thing for me. I need you to write ImpressFlow AI, a software to make presentations for me so that I can make presentations faster and spend less time on slide layout and things that I personally don't enjoy doing. You can go all the way up to there and then just, you know, let it be the thing that helps you ship. Find your level of comfort, work with it there, and find the right tool that works with you on the way there. Because all the features, all the audience, all the outreach, all the everything you can do, if AI makes everything transparent, right? Whether I'm using HeyReacher's API or Phantom Buster's API or scratch build my own one for some insane reason, it all goes away, right? My moat isn't my software. My moat is that people trust me to solve a problem for them. And if you communicate regularly on who you are, what you're doing, why it's special, why they should trust you, then you will get users regardless of the software itself. And then finally, data overwhelms you. Control it, harness it, use it to work with AI and use the AI to feed your decision-making. But in the end, you must be the human in the loop. You cannot abdicate your responsibilities to AI. I have seen people who are doing that right now. Uh, there's a dude who runs polis.ai, I think is what it is. He's got like 100 businesses all built entirely by AI and all of them are, they're super mid. Right? Will they stand out in a world where anybody can write features? No. What's his specialty? I don't know. He built an AI that builds AI things. If I wanted to hire someone to build my AI things, maybe he'd be pretty cool. But would I hire him to build my landscaping business? No. Right? Find out what works as you build your moat. Do more of the thing that works. As you build your moat, then your business will be able to grow and scale because you know the answer to the question that worked. Now, I promised you downloads. Here's the first one. If you want this presentation, it is available on ImpressFlow. Go get it. It's live right now. You can grab the image with your phone for a QR code, or you can type in the link and go. But more importantly, there's the skills. It's on Gumroad. The link is long enough where it's awkward to try and type in But go scan it or wait and Daniel will send you right after this the link to both the ImpressFlow presentation and the presentation on all the downloads for your skills on Gumroad so you can grab them. I would note I've had something like 250 people now, I think, grabbing these downloads from these workshops and finding them useful. They are completely free. They are something I hope you'll be able to use as a founder. And that's it. I think now your call to action, what you should do. First up, find a dataset that doesn't have any PII, run it in Claude and ask a big question. Then start to combine your sources. What story can you tell and what can you understand more together? And then finally, don't just do this, do this weekly. Ask your AI, how can you automate this to run it every week? Once the AI is pulling the reporting and it's showing you what you want to see every time, then you ask the AI, great, make this an every Monday or an every Friday thing. I would note right now in my company, I don't even plan my days anymore. I ask my AI on my tasks that are important in Notion, pull them, estimate how much work they are. Put them in my calendar, and then at 5:45 in the morning when I just woke up and haven't even yet gotten any coffee, post my daily scrum and fill out my calendar for me so I know what I have to do. That's a reiterative loop where every single weekday my day is planned for me based off of what I have told AI matters to me. But my reporting also runs once a week for every client internally. We see all this info, right? By making these automated, It stops being, oh, I gotta go run reporting. Instead it is, oh wow, here's data that I now learned a new thing about this client and I can make a better decision for them. You can do the same thing and you should do the same thing. I'm encouraging you to do it. Now, CrowdTamer, we build content engines. This is my pitch slide. If you need someone to build it for you, phenomenal. I'm not trying to sell you that. All I want you to do is figure out data and reporting. And how to use AI to understand your business better. All of you are here because you either are a founder or you want to be a founder. Go build something cool and tell me what you did.

Awesome. There you have it, folks. Trevor, thank you so much for another really great presentation. Everyone, if you like to see— I don't even need to say it, you already see the emojis on the screen. Awesome, awesome. I'm so glad, uh, everybody loved this, uh, but we do have a bit of time for questions, and I do some— these do see some in the chat. So if you have more questions, please feel free to add them, and I will quickly get to them. The, the big one, Trevor, was the concern about safety and security with something like cloud versus like your local hosted locally, and I think you did answer that. There was a really good follow-up here from Matthew, and he asked, what local AI models do you like or would you recommend?

I'm currently running, and the names of open models are all of them atrocious. Bear with me as I try to remember the specific model names here. The one I'm currently running most is Qwen-3, 30B MOE. Yeah, that trips off the tongue, doesn't it? They just rolled out this week Qwen 3.5 MOE. So I should— let me explain. MOE is mixture of experts. And what happens when you run an MOE model— I'm going way into the weeds here— when you run an MOE model, a 30 billion parameter model is going to tend to take up like 50 gigs of RAM. It's going to be huge. But when you run a Mixture of Experts model, it only uses like 7 to 12 gigs of RAM because all 30 billion of the parameters aren't active at any given time. It's instead thrashing on and off your hard drive to load what's relevant for the thing you're talking about right now. So I'm using that one. I'm using Qwen 3 Coder as well. They just rolled out Qwen 3.5 Mixture of Experts, which I think is like a 27 billion parameter model. Which I haven't tried yet, and a QWAN 3.5 coder. Everyone is saying these are both phenomenal models. I haven't used them yet because I haven't had time. But like, this is my vacation week. This is the last call I give before I hop in a car and drive to a water park in Toronto and attempt not to drown my children. Although depending on how I'm feeling any given day, I might try harder or not. So like, there's always new stuff developing right now, right? The models I'm using, QAN is quite good. It runs locally. It is Chinese made, but it doesn't go anywhere outside of my network. So I'm not concerned that they're spying on me. Mistral makes a decent writing model. Beyond that, like, KIMI 2.5 is a pretty good model. But unless you have half a terabyte of RAM just lying around, it's pretty hard to run. So I would like, for right now, for most folks using sensible machines, a QWAN model that's taking up 7 to 20 gigs of RAM will probably do most anything you want. GLM is also out there, but I haven't tried them, so they might be good too.

Okay, so for some of the folks who are completely new to this, right? So they're not really, they haven't started tinkering with any of this. So we do have a really good question here. You, at the start of your presentation, you mentioned interacting with Notion through an API. Could you be like a little more specific on what do you mean? What exactly is an API and how would you exactly interact?

Sure. So API is old-fashioned technology. It's Application Programming Interface. And for most modern web apps developed these days, What they actually have is a presentation layer, which is the UI bits that you click, and an application layer with a database that does stuff. And most of them these days, the application layer is queried and queries back and forth between storage through an application programming interface. So they have an internal API that powers almost the entire service. And then some companies, more and more these days, make the API external where somebody else can query the software directly through the programming and not through clicking on the buttons. Now, AI has limited context window. It only has so much data it can hold in its little brain at a time. I mentioned that Sonnet and Haiku and Opus right now have a 200,000 token context window, which is roughly 100,000 words. They are shifting it soon to be a million token context window, which is obviously roughly a half a million words, sorry, 500,000 words. And what that means though is it can only pay attention to so much stuff at a time. And if you're asking it to navigate its mouse around a screen and click, that's very expensive because this is a large language model. It's optimized for words and you just ask it to look at things and understand them and navigate and click. And this is not what it's good at. But an application programming interface is all words. They're programming words, but they're words. So it's way more efficient. So when you use an API, you're given usually a long password to connect to the API, and then your LLM can connect directly to their software and go back and forth and do stuff. It ends up being the most efficient way to transact with most softwares out there. And if you want to do, uh, what do I have I can show you here real quick? Um, I've got running right now. I use Obsidian. That's the text editor thing I've been showing you, which is a fair bit like Notion these days, but it uses, uh, files only stored on my local device. And they're all Markdown files. So the text is all like pretty, pretty easy to see, right? Like this is not hard for a human to read. It's also not hard for a language model to read. But what I can do is I'll go here, for example, I've now got a dashboard, which is all the projects I want to work on. And then every one of these is a Kanban file. So I can set these up Where's a— yeah, here we go. There's Impressiflow, which should be a Markdown file. This is not correctly formatted. I'm going to find a different one. We have here, there we go. Or better yet, we'll do one that's actually open. This one. When I make a change here, it's going to communicate back to Claude through the API and say, hey, Uh, that has now fired off a worker in like 30 seconds that'll start to work on this bug. It'll say, great, what's the problem, how do I fix it? It'll begin to churn through solving the problem, right? Something that simple can fire off a really complex chain of events, and APIs make that possible directly through programming language. That was a explain it like I'm 10 version of APIs. Hope that answered your questions.

Uh, no, that was super helpful. Uh, I think just kind of going back to analyzing data, right, I think one of the really important things is, uh, where AI kind of struggles is the context, right? And I don't mean like the context window, but the context of your actual business. Right, so it can only act on what you give it. So Trevor, if you have maybe any tips or advice on how exactly would you give AI sufficient data to help you make good decisions?

So a lot of this sits in skills, right? I mentioned your AI has two types of knowledge. It has transient knowledge, which is the chat that you're in right now. And then one of the big changes in AI in the last 3, 4 months is the existence of persistent knowledge, right? Your skills that you bake in, and also memory beyond that. If you work in the browser with OpenAI or with Claude for a while, it's going to start to remember who your name is, what your job is, all the things that you've told it, it will begin to retain. But specifics around like, hey, I always want you to design images in this style, that is better defined in a skill because that's easy to switch and change and customize. You've got these two different types of knowledge, right? Persistent and transient. The transient knowledge is the working state right now, what it knows as it's tackling a new thing, and it can pull in that permanent knowledge at any time. So when I mentioned, right, your 3 things to get you AI native, and one of them is you can control the context window. You have transient knowledge, in this case is actually these CSVs, the comma-separated value files you're downloading from Google Search Console and Ads and all these places. That is something that is, you don't want to remember forever. You want to use right now and then throw away. The permanent knowledge is how do I analyze this? What is it that you want to see come out from it? And when you codify that, it's just like how you would tell a new hire, right? Hey, we always want to see our weekly reporting dashboard with this page on front and then these pages behind in case we want to go dive deeper. That's something you would tell a new hire. It's also something you would codify into a Markdown file and make a skill out of, right? Because this is something any human can— when I have whatever, I just opened up a new random skill. When I have this setup here around, like I get every day a message sent to my inbox about what's new in the industry related to the models of AI that I'm using and why I should pay attention to. Cool. It goes through all this. It's AI written, but I can modify it. If I go, oh, I want to see a change here, I can type it in or I can dictate it in and say, actually, there's a new thing I want you to pay attention to and have it show up through dictation or however you want to modify it. These are human touchable. But they're permanent. That's one of the other big unlocks of the last 6 months is it used to be you had, and you still do in some cases use MCP, Model Context Protocol, which is basically a way to tell an LLM how to use an API to achieve a result. Now, I'm not going to show you any of my API-based skills because they contain my actual API keys and I don't want you seeing them. But instead of a model context protocol file, which is not very human readable, I can define in this, hey, use this API key, connect to this database, and then do these things. And the AI will follow that like in a model context protocol file, but it'll do what I tell it to. It's easier for me to edit and change. So when you're thinking about how to get ahead, how to make use of these different types of knowledge for your AI. Like, things came together in a rough in early December where we had agentic capabilities with OpenClaw. And OpenClaw only does 3 things. OpenClaw has, uh, scheduling, it has monitoring, and scheduling has memory and dispatch. Right? Those three things. And that's not a unique problem to solve, but once you combine that with, now I've got skill files, now I've got more capable models that can think better and retain the knowledge given to them in the file, and you begin to lump all of this stuff together, the pace of acceleration has accelerated. So. Your bit to remember is AI is iterative. Every time you use it, it's the worst it'll ever be. That's not quite true, but it's close enough. And you control the context window. When you control the context window, your outputs are better. And finally, AI makes software transparent. Whatever you want to use, if you use it through AI, all of the problems around how to use it are taken away from you.

Awesome. Super great answer. So we have a couple of minutes. I want to ask like 2 questions and then we can get into the networking lounge. And folks, you will receive a recording in the next 48 hours, so you don't have to worry about that. We'll also have all the resources that Trevor shared, so no need to save links or, you know, rush through the chat. Trevor, you mentioned skills. For the folks here, you can actually create skills on your own as well for your own context and the things that you're working on. Trevor, if you have maybe some tips on what's a good way to create skills and how do you think about it? Because it is like a specialized tool that can enhance whatever it is that you're doing. But like, you can't just ask it to create like whatever XYZ skill, right? There needs to be some thought behind it.

So yeah, you need to give it some definition. But two things. One is, I believe Anthropic has released a skill called the skill creation skill, which has the— Anthropic then has the AI Claude will ask you questions and the feedback you give Claude will determine what it does. For example, let me do this real quick and I will show you. It's always better than just trying to explain in the air. Here, enjoy my ludicrous pink window. I've talked about this before. I run Claude with dangerously skip permissions turned on. It is, as the name might clue you in, dangerous. And if you're going to do this, one, only do it when you're sure, and two, probably do it inside of a Docker container, a VM, or anything that's not your actual working computer. So that said, I get this live and I can say, hey, Claude, I want to make a new skill about email marketing. And it's going to go, cool, I need to know more, right? Or, oh, Claude's having outage issues right now, as I've already mentioned, so it may not actually connect just by having used it all day today. Oh, I need to log in. Interesting. Okay. Uh, Claudia Carlos subscription. Give me a second to authorize. I'm guessing they forced a cookie refresh on everybody. Yes, fine, go away. Okay, so try this again. There we go. It's going to ask some questions and be like, that's not enough for me to build a skill with, right? And it's going to fire off, I have installed previously a brainstorming skill. So it's going to say, cool, let's come up with an idea together, right? Doing it that way. Will give you a structured output, but you can come into this without giving it a perfect prompt. And when you have the AI, I've mentioned AI is iterative, right? Here is an example. A year ago, if I asked AI to write me up a bullet point list of how to make good email marketing and only said that, it would do it, and the result would be crap. Instead, this is saying, hey, that's not enough information to make a good decision. Let me, let me ask some questions to get better at it. That whenever you make a skill, I saw someone asking, is there an email marketing skill in my pack? I don't believe there is, but you could take any of the skills that is in the pack, are in the pack, any of the skills that are in the pack and say, based off of this skill, I want you to make an email marketing skill, but there's a couple of questions you're going to have to ask me first. Go off and make it, right? Treat AI like a partner, not like just a chatbot. Your results get better.

Awesome. And for my final question, because I know we, we've had a ton of people who joined this workshop, and folks, you can actually watch the replays, uh, for the other two if you want to. Trevor, what would you recommend? What's the best way to get started, right? This is the first time you're dipping your toes into AI. Maybe like the top 2 or 3 things to do, like right now.

Sure. So I mentioned my brother this morning. He's currently in Italy. So at 6 in the morning, I was up talking to him about AI. And he's at the very beginning stages, right? I asked him, where are you? And he went, I just, I just installed Claude Desktop today. So he's as early as you get, right? And I went, cool. What do you want? And there's a very particular technical reason why he installed Claude Code. So good for him. And there's an error that's fired in his company. And when it fires off, it used to be he'd have to wake up at 5 in the morning or 2 in the morning or whenever and go fix it. And he's like, well, but I, I get called. I don't have proactive errors coming my way. It's someone says, hey, I tried to do thing and software is broken. So you have to fix it right now. And he went, so I want to have AI help me write software that checks every 15 minutes. And then if it's down, it tells me. And I went, that makes perfect sense for a first AI project, right? This is, you build your ladder up, you begin with just It is a research tool, then it's an optimization tool, then it's a prototyping tool, then finally it becomes a builder, and then lastly it becomes an agent, right? That's the stages up I talked about in my first presentation. He's down at the, just the beginning researcher, right? But that's a phenomenal thing to do. And for a first project, it's a great one to tackle because AI is going to help you write code. AI is going to teach you how it thinks, for values of thinks, how it works. Cool. But here's my question. For you, younger brother, which is why don't you have AI fix the problem instead of email you? And he went, but then what do I do? I'm like, well, something better than wake up at 5 in the morning and fix something because somebody else broke it. Like, so you start with the thing you're comfortable with, which could be email me if it goes down. That's already proactive. Your boss will love you. But then make it where— don't just email me that it's down. Email me with the error code. Why is it down? Now that you've got me the error code coming to me, you've gone from research up into optimizing. I can go, oh, I know what the fix is for that. I can already log in being aware of what needs to be changed and change it. Then you can go to prototyping. Hey, instead of just giving me the error code, Tell me how you would fix it. Don't write any code yet. I don't want you to muck around in my software. Just tell me how you would fix the problem. And then I'll go, no, that's not right. And we'll write a skill together over a couple iterations on how to do it right. And then you go, cool. Now I want you to automatically fix the problem if you are 100% confident you know how to fix it. Otherwise, email me. That's builder level. Finally, agent level is I want you to call me on the phone. Day or night, through voice-to-speech and speech-to-text— there's text-to-speech and speech-to-text, there we go— call me up and say, hey, it's broken. I think this is how I fix it. Is it right? And you can say, yeah, that's good. And then go back to sleep. Those— that right there is a great model of the 5 stages of AI-native work and how you make your way up of it. But don't try to start at the top. One, that sounds like science fiction, but two, because you'll be super uncomfortable getting there. And honestly, if you didn't follow all the steps to get up there, it'll ruin everything in production for a Fortune 500 company and make everybody really mad at you. So work your way up. Start with just, how does AI help me do a better job? And eventually, how does AI take the repeatable work away from me so I can focus on stuff that AI can't do yet?

Awesome, awesome. There you have it, folks. You have an entire framework that you can use to start working with AI. Trevor, it's been a pleasure. Thank you so much again for being a part of our workshop series. Folks, if you don't know, we actually did this last year or two, and you can actually catch those videos on our YouTube channel as well. Manu has mentioned all the links in the chat, so do check them out. We will be open for networking for as long as you're around. So I do hope that you all stick around, talk to founders from across the world. Uh, Trevor, before we close out, any final words here?

I think what I said at the very beginning of the first workshop, right, the cost of effort is dropping to zero. That's terrifying. Society is going to have to rearrange around this, but also right now and for the next couple of years, there has never been more opportunity. All of you here listening, all of you have built, are building, want to build. My final words to you are go and build. It's never been easier, and you've got a window right now before it almost doesn't matter what you build because building becomes not just free but also instant. Now is your time to go build something to shape what it's going to be like in 2035. That's hopefully exhilarating. It's certainly terrifying. But I hope it'll get you up out of your seat and making something cool.

Okay, folks, so make something cool. We will see you at our next event. I hope you all have an awesome day ahead. Bye.

Take care, everybody. Bye.