Adam Sand demonstrates an AI workflow for creating a roofing brand film and running it as a Facebook ad. He uses Claude to extract Soul IDs from Higgsfield, builds JSON specs for character consistency, spawns three AI agents in parallel to generate scenes, and launches ads with a Meta token. An AutoResearch loop then tests new copy thousands of times faster than traditional agencies.
Chapters
All right, quick little check to see if we are live. That is the most important thing. So one second. This is gonna be an exciting one, guys. You're definitely not gonna want to miss this one. Let's just make sure that we are in fact live and working because I told you I am, I'm not great at live streams. This is the first one and we are live. Perfect. So we are good to go. All right. So what are we doing today? This is going to be super exciting today. We are— they released the MCP and the API for Meta ads, meaning that you can now connect your Meta ads to an agent to run ads, test ads, improve ads, and actually make your ads better through something like the Andrej Karpathy Auto Research Loop. All completely from an agent. So you can have AI run your Meta ads for you. So you can take the equivalent of a 200 IQ person that knows everything that was ever made and have it run your Facebook ads. I'm going to show you everything, how to connect it to your agent. I'm going to talk you through the Meta business ad creation process, and we are also going to generate a Super Bowl level ad today. Why is my hat all screwed up? We're going to go through a Super Bowl level ad creation today with Higgsfield, and I'm not shilling for Higgsfield. I don't get any money for this. I do get money, however, for Roof Claw. And so if you want, I'm going to be showing you how to do this through connecting a Roof Claw instance to the Meta Business Ads and connecting it to Higgsfield. And we're going to generate a Super Bowl ad here live. Again, I do these live streams at 9 o'clock. This is my second live stream in history. I fully expect it to suck. And so you probably shouldn't watch it. But if you want to watch me bash my way through this, then we will do— we will get through this together. And I'm excited to show it to you because honestly, I'm pretty excited about the idea of running Facebook ads from an agent. Because here's the thing, the Facebook ad platform changes all the time. The buttons move around. Like, I— my whole business started because I did Facebook ads for my roofing company. That's how I got into this whole situation of running, like, Roofing Business Partners, a company, was because I got famous for doing Facebook ads in 2016. and I helped a bunch of roofers with that back then, and then that's transformed into everything that this is today. And so over the years, when it went from Power Editor to the Ads Manager, I started to get confused inside the Facebook ad platform. I know what needs to exist. It's just that I don't necessarily always know where the buttons are inside it. So the ability to use the API to run Facebook ads from an agent while taking advantage of a 200 IQ that can read at 150 words per second is amazing. And so there's a couple things that I want to show you today. One of those things I'm going to show you is we're going to be talking about how to actually connect your roofing, uh, your, your roofing company's ads to an Open Claw agent. So that's actually working through— you can see here I tested it, I managed to max out the, uh, the API credits on here. We're going to be using this ad for today to actually run the— to actually post the ads to the Facebook Ads Manager so So we're not gonna be doing the whole thing inside here and then run into this rate limit. That was me testing it earlier, but that is all I have truly tested. The other thing that we are gonna do is we are going to, we're going to be talking about a tool called Higgsfield. And again, no sponsored, I'm not sponsored by them, I don't make any money, but I've always wanted to get into this and actually use AI to do it. And now I finally have the opportunity to do it. And that's this tool Higgsfield, which is super cool. It allows you to generate videos. It's truly like giving a video producer everything they need. This is the extent of what I've done inside Higgsfield. This is all I've done is I've created 3 pictures with Higgsfield. That's it. But there's so much more to this tool. Like you can actually create like consistent characters. You can create all kinds of like, like you can use these incredible video models. Everything you see right now, if it looks good with AI, which I don't know how, how many things you've seen. There are like full cinematic videos that convince you that Elon Musk and Neil Patrick Harris are running, uh, like an end of the world video. Like, if you think that the AI slop is the equivalent of all you have, like all AI video has to offer, you're crazy. And so I'm going to walk you through creating a Super Bowl ad. Now, if you want to take advantage of all this, if you want to actually walk away and be able to use it. Um, what I will tell you is that I am putting everything that I'm sharing today— everything is going to be completely free. You will be able to point your Roof Claw or your agent at this GitHub repo, um, and you will be able to completely walk yourself through. So as I go, you will see me posting commits to this thing. So you can see I already did one pull request here. It is closed, where we added the Super Bowl ad creative process for steps 1 and 2. That's because I've done— I have my own video that I've been dying to make for years, and a video that I've always wanted to make, something for my business. And so today is my excuse to do that with you guys. So I'm actually going to go through it. So if you guys are here and you guys are digging this, leave me a comment, something like that. Let me know you guys are here and in the room, and if this is something you guys want to see. I don't do live streams all the time. I'd love to know like who's here, where you're com— where you're coming from, what what you want to do. But I want to show you one last thing before we dive into it, give everybody a chance to kind of just get a picture for what we're doing. I have always wanted to create this, this video for literally years. I've always— it's called Quiet the Doubters: An Ode to the Roofer. And there's an inspirational video that I saw that was for a software called Infusionsoft. And this software, legit, like like, you know, boys, we're all men here, but like legit brought a tear to my eye years ago because it talked about how hard entrepreneurship was. And I was like, man, why is this video so good? And so I watched this other video called Story Beats. It's called like, basically it's called Save the Cat, and it's like the Save the Cat beat sheet. It's how every single video you've ever watched, from Donkey Kong to, um, to Shrek to Die Hard, is the same movie. Right? It's the same movie with the same beats. You set the scene, you introduce the character, the character has a problem, the problem is internal, external, and philosophical, or all three. And then it introduces— so they're the hero, it introduces a problem, then the guide shows up, the guide has a plan, the plan has a bunch of like fun and games, and then there's this climactic challenge, and then either the world changes around the character Or the character, character changes to the world. That is every movie ever made, the Save the Cat beat sheet. So I was like, I wanted to work on this. Now, if you look in here, if I go in here and look at the audit version history, okay, so if you enter the version history, I first created this document August 30th, 2023 is when I started basically breaking down this video into what I thought I could turn this into a video about roofers. And now I could do this for my business. And so today I went through a process of turning this into a, um, an actual storyboard. And so what you first need to know is that, um, and I'll put this in the comments, is that if you guys want this thing, right, you can come into Roof Claw Meta Ads Agent and you can point your Roof Claw at this link that I'll post as a comment. You can point your agent at this. It's going to get better, but it's going to get better as we go. But this is the thing that I've always wanted to create. And so to generate this, I've been putting all this together for you guys. So this— the best part is it's very, very, very, very simple to actually connect your Roof Claw to a, uh, like, to Facebook Meta ads. I want to show you the exact sequence that I went through today. All of it's in the GitHub repo, so you don't have to figure it out like I did. You literally just have to plug it in and say, walk me through it. And there's a file in there called Launchpad, and it will walk you through it. And so I talked to my Roof Claws through Telegram, right? And so you can see way up here, right? See, that was where I ran into the rate limit. And so if I come all the way up here, right, I said, hey, I want to I want to connect. I said I was going to be doing this demo today, and so I pulled the blog, got everything in, and okay, so we've made a decision. Are we going to go through the MCP or go through the CLI? The CLI has more power than the MCP. And so I said, hey, we're going to probably do this with Vito on a demo tonight, so I want to make sure I don't look like an asshat, right, and work out all the kinks. And so, um, we worked through, we tested the MCP And so if I go over here and I show you guys in, you're going to go, if you want to get the, if you want to kind of follow along or get at least the process started, you can go up here and I'm going to go to Chrome tab. We're going to go, we'll go to this one. Okay, so this is where I hit the rate limit. So you go to developers.facebook.com. Okay, so you're gonna go to developers.facebook.com and you're gonna go into My Apps. And then when you get into my apps, and I got to make sure I don't cook myself here, but you get into— you're going to create an app. So you go through this process of like, it's very, very straightforward. You get in here, you go to create an app, you're going to go in here, you're going to make a name. So I'll say like Vito-RoofClaw. Okay, and then you're going to tell your use cases. The crazy thing is, is look at everything it has. Create and manage ads with the marketing API. Yep, want to do that. Measure ad performance data. With the Marketing API. Absolutely. Capture and manage ad leads with the Marketing API. Yep. Create and manage app ads. Well, I don't really run an app, so that's okay. Access the Threads API. That's, uh, Meta's equivalent of, uh, like x.com. Don't really care. Advertise your app. Okay, don't really need to do that. I don't need to request or authenticate people. I don't make a game. Manage products with the catalog. I mean, I guess I could maybe put Roof Claw into the catalog. People could buy that. Engage with customers on Meta, from Messenger, from Meta. So yes, my, my Roof Claw, I can create a subversion rule set. I can basically say, hey, when you're communicating through this channel, um, the Messenger, that is to customers, so don't talk about my personal shit, right? This is what's possible. And then, um, I don't really, you know, don't really have a lot of fundraisers. Message, manage messaging content on Instagram, absolutely. Access the live video API. Yep. If you're a roofer and you're not recording Facebook Lives, you're crazy. Put your hands up if you record Lives for your roofing company. Super, super viable. Embed Facebook, Instagram, and Threads content in other websites. I mean, maybe. I don't think so. Manage everything on your page. Yep. Right. Connect with customers through WhatsApp. I'm not in Mexico. Don't need to do that. So then you just click next. All right. And then you can link it to a business. So I've already done the business verification over here on my business. So then you click next, and then there's no requirements identified because I've already verified. If I go previous and I say something like this roofing company, if I go next and then maybe next, I actually don't know, it's probably going to have to give me some— I probably shouldn't do this, this is a roofing company that I used to run ads for way back in the day. But then you connect it, and then what's going to happen then is you're going to have an app. So once you have your app in the system right, then it'll look like this. It'll be in a demonstrator mode, so it'll be unpublished. You're going to have to give it your privacy page URL, and you might need to use ChatGPT to edit your privacy page to match the requirements. So basically say if people request for you to delete their information, you would delete their information, and you have to have a place and a way to do that. You probably already have a privacy page if you've ever ran Facebook ads before, so because it requires it there. So if you go to your roofingcompany.com domain and and ask for a— if you ask for a— what is it? Oh, you'll go roofing— you're roofingcompany.com/privacy. That's usually where it is. You're going to put your privacy statement in there and then you're going to hit publish. Okay. And so it'll ask you to go through the process. You're just going to have to follow. Again, if you have a Roof Claw, it'll walk you through this, right? It'll walk you through the process. Then you're just going to go through the screens and talk to it. My GitHub repo that I posted for you, it will walk you through it. It'll walk you through this whole process. This is not the part that I want to spend a whole bunch of time sharing because it's literally just setup. We don't need what we call setup porn. I actually want to talk about why you would do this. If you're here and you're late with me, you know I like the zonic or the zin, so zin and lock in because we are going to walk into the storyboard and Super Bowl ad creation process, so rock with me. All right, so when you guys are creating ads, the reason that AI videos suck is because AI video is like plastic surgery, okay? You only notice it when it's bad. AI video is like plastic surgery. You only notice it when it's bad. When it's good, you don't know the difference. There's a guy that sent— named Matthew Berman. He sends me a— he sends me an email like every other day, his newsletter, and in that email He has a picture or a video and he says, is this AI? And you watch it or look at it and then you click a button, AI or real. And then every day he sends you whether it was real or AI. And it's like down to 80, 80% of the time people guess wrong. And so AI video and AI in general is like plastic surgery. You only notice it when it's bad. So the goal is how do you be good, right? Well, like anything, like roofing, the magic is in the preparation. Okay, and so in that, in that, that, uh, GitHub repo where you guys can point your Roof Claw— if you don't have a Roof Claw, go to roofclaw.com, buy it. We build the best open cloud instances for roofers. We will train you how to do it. It's safe, secure, shipped to your door, trained on your business, and then we, we spend hours with you training it, and we support you, and we even build a little mechanic and agent into it so that when it breaks, it's fixed, it's fixable. But in here Once we— so again, you can see here, walk me through the process, right? And here's a little secret tip: when you're working with agents and you want to send passwords, use pwpusher.com. Again, not an affiliate, not related, no, get— don't get any money. pwpusher.com, it's free to use, but you can send secure tokens to your agents. And then it walked me through and we tested the API. And so I put all the API endpoints in there so you understand what's, what's all possible. And what's, what's not possible if you go through the MCP. And so I tested a creative upload, we ran into a flaw, that's where we went and fixed it. And so it's published and now we have full chain complete. So then we tested it out again and I said, okay, this is what we're going to do, this is what we're going to show you guys today, all right, is how to come up with a story for an ad, write the storyboard, generate the images for the storyboard, turn the images into a video, tune up the video, post the video to a Facebook page and run an ad from the video posted. And so this is going to take a while, and I'm going to walk you guys through the process so you guys can quit and come back and watch this later. Um, but you— I want to walk you guys through the process. So I've, I've gone through all this, got it all posted, okay? And then we tested this out, and now we're at the stage of— all I've done is with my other Roof Claw, Vito, right? Vito Clawlioni. He's the Clawd Father. He's the one who helps me sell manage, build the roof cloth side of the business, we have generated a story. We've generated a storyboard which I have now on my desktop and we're going to give that to Claudio, who is the one who's actually connected to the MCP. I want to walk you guys through this. Which is better? Is it this view? Is it this view? What's better? Let me know in the chat. Is that one better? I don't know, maybe this one. So going in here, we're gonna say, okay, we're gonna make the Super Bowl ad. So he's committed this new skill, and now we're gonna go— oh wait, did he commit it? Push someone, probably out of something, get up. Yeah, okay. So I'm gonna go /new to create a new session. Okay, so wiping its brain. The, the crazy thing about internet— or sorry, about AI— is that you don't have to be a good leader, you have to be a good trainer. You don't have to be a good leader, you have to be a good trainer. It's the perfect employee. You can talk to it for 2 minutes or 2 hours. You can talk to it at 2 in the morning or 2 in the afternoon, and you can tell it to do something twice and it doesn't cry. But you have to know how to train it. And it's like a Roof Claw is essentially— it has a soul.markdown file that teaches it who it is. It has an identity.markdown file which gives its name. It's got a users.md file, so it knows you and knows your team, knows who it's allowed to who's allowed to use it, who they're— to what degree they can share with it. And then you're going to have things like your tools.md file, which tells it what tools it has access to. And so we made sure that before I quit that last session, that we committed Higgsfield and Meta Ads and this GitHub repo so that it understood what it is, because it's got a 200 IQ but it wakes up every day with amnesia. So we started a new session. I'm going to go think because I want to make sure I set a high thinking style. We're going to set adaptive. We're going to say it can adapt to thinking. We're probably going to burn some serious amount of tokens tonight. Now we're going to, um, make sure which model are we on. So we're going to say which model, okay? And we're probably going to say, okay, we're on Opus, that's good. So we're going to stay there. So I got Opus 4.7, I got adaptive thinking. So we're going to tell it what it's going to do. I like to talk because who likes to type anymore? So I'm going to use my Whisper Flow to do that. All right, so last session we worked on the, um, Meta Ads agent framework for Roof Claw, and I had Vito develop a storyboard that I want to send to you. I'm also going to send to you the Higgsfield creative process research that I put together, and you also saved a transcript from a video. The whole goal for you is going to be the— I want you to use the MCP to start the storyboard creation process and actually start to design in accordance with the Higgsfield best practices and walk me through creating a video for our customers, which is called Quiet the Doubters: An Ode to the Roofer. This is a video I've wanted to make since 2023, and you are doing this live on a demo on, uh, in front of our, our customers, our friends and prospects. So don't, uh, well, not that if I tell you to not make mistakes, you're— I'm not— I'm going to delete that part. Friends and prospects. Um, all right, so now I'm going to say, okay, here we're going to give it— we're going to give it some files. So again, this thing is sitting on a MacBook, a Mac mini over here. I have, I have a Mac Air, which is Vito's machine. I have a Mac mini, which is Claudio's machine. And if you see behind me, there's a whole bunch of laptops and stuff like that behind me because we're in my home office and I build a ton of these for people. So I'm gonna give the Ode to the Roofer storyboard, and I'm gonna give the, from Turning Your Video Idea into a Visual Ad. Okay. And so we're gonna give it this. And we're going to give it Higgsfield best practices. All right, so I'm going to let it research these documents. These are all in Markdown files. So you'll notice I use Markdown. Why not PDF? The reason I don't use PDF is because a PDF consumes 40x the tokens of a Markdown file. So that if you only have, say, 200,000 or a million token context limit, well, all of your pre-configuration files, all your conversation, all the reasoning that you're going to do to work through it You don't want to burn through that making it read a PDF when it doesn't need a PDF to understand the content. It doesn't need your paragraph structure. It doesn't need your fancy formatting. It just needs the words. A hashtag tells it's a headline. Okay, so now I know we have a heading. Two hashtags is a subheading. It knows it has a subheading. It can understand the paragraph structure with one character instead of a bunch of fancy designs. So we use Markdown files instead of PDF. That's your writer downer of the day. Alright, so I'm gonna talk to a little bit more. I've attached 3 files, an ode to the roofer, the storyboard. That's what Vito came up with. I've also got the from, from Super Bowl ad creation process to the Higgsfield creative process. I've got that file in there to walk you through how we should process this with Higgsfield. You and I have never done this before, so you're gonna have to walk me through it. And then I also have the Higgsfield best practices markdown file, which should give you an idea as to how you are going to use the MCP. I would like to use cDance, the video. I give you the creative freedom to use whatever you want for images. Follow the best practices from the transcript of the YouTube video I gave you before and this and come up with a plan and show that plan to me here so that we can show it to our audience. All right, so now we're gonna let it work. So while we're waiting for that to produce, walk me— like, say hello. Let me know where you're coming from. I see Jack here in the chat. Jack, it's good to see you. Jason, I, uh, I, I love having you here, man. It was good chatting with you earlier. This is like, this is the, the, a lot of the prototyping of agents that goes into a much bigger project that I'm working on. And that's why I give this stuff away for free. I want people out there to use it. I don't want to gatekeep this stuff. I want you guys to be able to learn. I want to know who my people are. And so I encourage you guys to, uh, to comment and engage and be a part of this stream because that's the only reason you do a live stream. Right? You do a live stream so the audience can be a part of it. And so I want to know who's here. I see your names in the chat, I see— or I see your faces, see your eyeballs. Would love to know who you are, where you're coming from, and, uh, what questions you guys might have. And so while we're doing— while we're letting this thing pound out, it's basically— well, no, we will communicate with what we've got. So what do we got it producing here? So the plan, Adam, ready for stream. Here's where we walk the audience through this. We're not going to produce all 15 panels of Vito's storyboard live. That's a 3-minute hero film. 25 generations, $60 in credit, hours of rendering time. I don't give a shit. I burn the tokens, okay? I am, I, I am all about burning tokens, okay? We— I measure my self-worth in tokens burned per day. Uh, we're going to get something better for a live demo, prove the pipeline end-to-end on the most important single beat in the film. The live demo plan, an ode to the roofer, what we will actually ship tonight. Okay, so Claudio is telling me what we're doing. What was your— the mirror shot, panel 1, opening image, lonely stuck, a lonely truck, Predawn. So maybe it would help you guys if you guys could actually see like what does the input for that look like. What did the, what did the input for that look like? So, uh, what's the best way to show you guys that? Probably in anti-gravity. Actually, it's a good lesson on Markdown. It's a good lesson on Markdown if we do it that way. So I'll stop screen sharing and I'm going to go and— Ode to the Roofer storyboard, show in Finder. And so again This was done with the help of AI. We're going to go over here, max out screen over there, and we're going to go share screen, window, bang, bang, bang. Markdown, this is Markdown, everybody. Markdown is 40x less tokens for your AI to consume. Yes, you won't like reading this this way. You're not going to enjoy reading this text. But what, what is Markdown? Markdown is a 25-year-old alternative to Microsoft Word, okay? Before we had visual computers, when computers looked like this, like when you looked at a computer and it looked, looked like this on a screen, well, back then you didn't have, um, you didn't— you couldn't see formatting on your screen. Like, there wasn't a word processor with formatting and fonts and italics and bold and all that kind of stuff. So you wrote a document like this, this little hashtag right here. Okay, um, up here, okay, this little hashtag, that there was saying this is a headline 1, 2 hashtags headline 3 or 2, this headline 3, then this meant it was bold. And like when you put a bracket, like a, what else, up and down line, a vertical line here, this meant like it was a table. So if I click right-click preview, this is what happened on the printer. The printer did this conversion for you, and this is why it consumes 40 times less tokens. Matt, it's good. Yes, Whisperflow for the win. I absolutely Matt, you and I got to get talking about CapOut and how we can connect Roof Claw to CapOut. That should be the next skill that we put into GitHub, is the official Roof Claw CapOut skill. We got to get that booked. In fact, Matt, um, Matt, I'm going to— while you guys look at this, I need to send you— send me a message on Facebook. We got to, um, I got a DM because I'm gonna— I'm gonna send you a calendar link. We gotta work this out. We gotta get together to build the official Roofer Capout skill. Go ahead and throw a link to Capout in the chat, let people know what that is. We are going to build the best integration to Roof Claw and Capout to save you guys time. So Matt, tell people what that is in the chat and throw a link to Capout in there. All right, so this was help designed by AI. So I used— I took that document, that one that I wrote in 2023, and I, and I did a research on Perplexity, and I said, hey, what is the Super Bowl ad creation ideation process And then I took that, gave it to Claudio or Vito. I took it to Vito. I said, hey, here's the Super Bowl ad creation process. You know Roof Claw, you know this business. I want you to position Roof Claw in as the guide, thinking about the story beats, right? As the guide that helps the hero, which is the roofer, and shows through this video like a Super Bowl ad that, that Roof Claw can be their guide to help them with a plan to get them away from their pain and towards success. And so it used the Super Bowl ad creation process to write this. And so it broke it down into like literally second-by-second beats, right? So it had a full storyboard. And now we have to turn this into the visuals. So that's what we're doing here today. And so when I come back over here to Claudio, that's what we're working on. So I will share screen and come back and we're going to go back to Telegram. I'll keep— we'll keep it going. So, so it's going to go, okay, 6-step live walkthrough. What the audience watches us do. Step 1, the translation, 5-minute walkthrough. Show Vito storyboard on screen. Hey, I already did that. Step 2, cast Marcus, character lock per Higgs field best practices. You lock the character before we lock the shot. Use Nanobanana 2 to generate a 4-angle character reference sheet from Marcus. So this is how we get character consistency. It's going to make it so that we have a character that has been viewed from all sides. So then when video tries to generate the character, we get consistent characters throughout. Because again, remember, AI video is like plastic surgery. You only notice it when it's bad. All right, we all know what we're talking about when we say that. So, um, we're going to cast Marcus, then we're going to generate the hero frame. Okay, so I'm going to say here, all right, so, all right, I've explained to the audience um, the Ode to the Roofer storyboard. And now I want you to start on our next step, which is going to be casting Marcus. And I want you to send me the images so that the team can see, that everybody here on the video can see the images that you generate. Okay, so again, this is the prep, this is the upfront. All right, this is the upfront. Matt, I got your message, I got your Calendly link. Right on, jam out. So, um, we want to cast our character because It's all in the preparation. If we get the characters developed up front where you can see it like this, and then you can see it like this, and then you can see it like this and like this, right? And you get the zoom-in shot and you get the eye color. Well, then we can start to make it so we understand. We can create a JSON. JSON is just a language that computers talk. We can generate a JSON representative example of what, um, Marcus looks like. And his character traits. He's got a, you know, like if you're doing me, you'd say he's got a crooked tooth, he's got a big nose, he'd say he's got beautiful eyes, you'd say he has a poorly grown beard, right? But he's still ruggedly handsome in a very Canadian kind of way. That would be the JSON of Adam Sand turned into it. You say he's got a hat from a church and it says Courage Crew on it, right? Like that, that would be the JSON representation, right, of the Adam Sand character. So that way, as you generate successive videos, you get the same character. And I, I have done all this kind of stuff in VO3 and Google Flow. I've done some of this stuff before, but it's always been like me in the screen. It's like, what do I mean by in the screen? Because you can use— again, you can use Higgsfield directly from the actual screen itself, right? Like, you can go into Higgsfield and you can click around in here and you're like, okay, I want to create an image, right? And you can come in here and you can design— you can create the image and it'll design it, right? But then you would go over here to video and you'd go like image— you'd add your image to it and you turn it into a video. You could use this software, but who likes software anymore? The whole idea is how can we make it so that we can work from one screen? Because here's the magic of what's going to happen after this is as I go, I'm going to document these as repeatable skills, repeatable skills, and then I can combine those repeatable skills into a directive which then can repeat each skill. And it's gonna— I'm gonna try and make it token efficient, try and create Python scripts to make it repeatable and duplicatable, work out the bugs. I'm gonna upload it to GitHub so you guys have access to it too. But then you can be like, hey, here's a storyboard, I want you to create this video. And this is the beauty of Roof Claw, is you have the heartbeat.md file, which means that on a chronological trigger you could say, hey, every day I want you to create a video, and then you could upload a storyboard every day, and then it'll keep its heartbeat going every 30 minutes. So you don't have to sit here and support it. It will just keep working through it and keep itself awake, and it will keep working through the steps of the directive to accomplish all the steps in the skill. And you have a single input. That's the whole magic of it. It's like an end-to-end workflow that keeps itself alive, but along the way, it's using its 200 IQ to make good decisions. And so fine, again, this is not a shill for Higgsfield. I'm not sponsored. I know there's a lot of guys like that on the internet, but we have our character, we have Marcus. Okay, so we have, we have Marcus here. It's a handsome looking fella. Okay, and now we have him from all angles. So we have some consistency. We understand he's got these rugged boots, you know, his color of his jeans. We understand his face from multiple angles. We understand how his neck wrinkles when it curls. We have a picture for his hair. So we should be able to get character consistency. All right, now if I go into Higgsfield, I can probably find it in there, and I can probably see like, okay, because it is actually in— again, we— who wants to use software these days? So if I come in here, image, do I have character? No. Where is— I don't even know where it puts it in here. Character in here? Nope. Where did— I don't even know where it put it in this thing. If I go Marketing Studio So put it in here. No, I don't even know where it is in the software. That's the crazy thing. That's the crazy thing is you don't even know where it is. It's just, it's in the software. You're using the power of the software and it's doing it all for me. Is this it here? Okay, so I found it in all assets. Perfect. And so you can see here, professional character reference sheet, 4 angles in 1 image. So now it's saved this, and now I could click animate here and I could go forward I don't want to use the software, I just want to work with my teammate and actually build out my skill so I don't have to do this, this kind of stuff all the time. I can just work with it to say I'm going to build this into my agent so I can work with my Roof Claw and I can always do this over and over again. So jump back over to Telegram. Who's digging this? You guys learning something today? All right, so job submitted, waiting for completion. So what did it say was step 2? Okay, so we've cast Marcus. All right, so Jeremy Renner energy, work jacket. That sheet comes with our consistency anchor. That requires 20 to 30 real photos and 10 minutes of training. We'll fake it cleanly with the reference sheet. I don't want to fake it cleanly. I want to do it the right way. I want to do everything the way we're supposed to be doing it. So if you need— if you're saying you need 20 to 30 shots of Marcus, then we should do that. So however you need to do that. If you need to work through multiple, like, multiple rounds of this, if you need to spin up like 4 sub-agents to go and create— 4 times 4 would be 16. So if you need like 4 agents that generate another casting of 4, so you can get like the zoom in, the eye color, you know, action shots, whatever you need to do to generate our character reference sheet and get these, uh, get these pictures, do that now. I don't want to fake it cleanly. I want to do it right because we're doing this live in front of people. We want to teach the best practices. We're not doing one of those YouTube videos where we fake it and then people walk away with the half-baked process. I got all kinds of time. Let's do it the right way. All right, so we're going to let that go on. So what do we got here in the chat? Jack, are you still around? I'm curious who's still here. We got some likes. And again, I am not a live streamer. That's why I do it at 9 o'clock at night. I fully expect this to suck. So I, I, I'm doing this because I actually want this project. Like, for me, I'm just doing this, I'm bringing you guys along for the ride. The only thing you guys are doing is like not letting me play my music. And when you guys— if you guys all end up dropping off, I'm just going to jam out to music on here and I'm going to finish this project because I'm excited about it. So if I come back over to my assets, I'm just going to check on the other screen to see if it's generating more. It, uh, not— it hasn't come through yet, but I'm gonna get better at, uh, switching the screens. And so here's share screen, come up here. There we go. I'm getting some comments back. Come in here. So like, this is just the test project that I did earlier. I was just like, hey, can you generate some images? And so I just want to make sure the MCP worked. And so it looks like we have, uh, Jack Reacher up on the roof here ready to beat up his neighbor. Um, I said that the— see, the one thing I want to make sure we do is that when we train our AI to do this, I'm going to have to train this guy on we don't use claw hammers on roofs. That's the thing. I'm gonna— I want to make sure that it gets into our skill that when, when Roof Claw generates videos, it doesn't generate videos with claw hammers or hard hats or any of the other common tells of AI video. That's something I promise I'm going to train into this skill before we're done tonight. All right, so coming over here. Did you generate the photos from different angles? Yes, so it is generating them from different angles. So you'll see here, um, you'll see here, Jack, that when I said, hey, I want you to, to cast Marcus, it used the Higgs Field Best Practices skill, right? So there's the— so I gave it two primary skills already. One skill was Higgs Field Best Practices, so it understands all the endpoints that exist on the MCP, Model Context Protocol. That's your trick here, is that Higgsfield has an MCP that you can connect to OpenClaw. And so you can say, hey, go, go connect to the MCP. And then I did a research project and gave it all the best practices. So I said, hey, let's take all of these YouTubers that are talking about how to use Higgsfield, took all these blogs and, you know, people who are selling info products and have downloadable freebies, and I shoved that all into my research project with RoofClaw. And I said, hey, turn this into a Higgsfield best practices thing. Take all the best stuff and put this into a mutually exclusive, collectively exhaustive document of all the best practices for Higgsfield. And I gave that to it as a skill. And then I said, hey, there is an ideation process. Like, why are Super Bowl ads so good? Why are they so good, right? I'm like, why is that? And so I said, well, I wanted to say, um, what is it that What is it that makes them good? Is that there's probably an ideation process. There's probably an upfront preparation process where they put a bunch of creative people in a room and come up with an idea. Now, I had an idea. I wanted my ode to the roofer, right? Like, I wanted that to be my, my video. So I had a bit of ideation already done, and then I said, okay, now how do you turn that idea into something that has all the emotional hooks, has all the bits of the story that has all the bits of his story that would actually take it and turn it into an actual Super Bowl storyline with all the, the, the, the kind of setting, the mood, and what are the facial expressions. I turned that into the storyboard, and then I gave it another skill, which is how do you turn the story into the visuals. And so that's where it said, hey, the first thing we need to do is casting, right? Hollywood movie, Hollywood or a Super Bowl level ad, we need to cast a great character. And so this is a casting a character. So, um, you said I want to try with a real person, would it work? Yes, absolutely would work. That is one of the very common use cases of Higgsfield is digital clones. A lot of people do that. So people do this all the time, whereas probably like AI Influencer, there's something, there's something in here I've seen. Again, I haven't used this. That's what I love about MCPs is that I can use stuff to say um, to say like, I don't want to actually, you know, uh, how do you say it, um, I don't want to have to make— like, I don't want to have to learn a whole software every time I want to do something. I just want the AI to use the MCP, take the front-end user interface out of it. I don't want to use the front-end interface. I want to just have it. And so when you have a character— character swap, face swap, AI styles, find your perfect look— like, in here you can generate your aesthetic. You can create you as a— as an— and— or you or someone else So you can choose your character, or you can take a real person and essentially create a digital clone of that person. The same thing as like HeyGen. HeyGen allows you to do that too, but this thing goes even further to say you could also have an influencer that's a crocodile if you wanted to. And so it really is— the limit is your own creativity and the preparation that you do. All right, so we're going to go back over to our assets and see what it's put together. Still put more together. So let's see what— let's see what our Roof Claw is saying. All right, so we're going to go here, stop screen. We're going to pop back over and see what Roof Claw's got to say. All right, so what do we got here? So, right call, doing it properly. So, the idea is the best practice from Higgsfield's own docs: 20 to 30 reference images, varied angles, expressions, lighting, and context, same character throughout. The more variety in the training set, the better Marcus generalizes when we drop him into truck cabs, kitchen tables, rooftops, and golden hour resolutions. The plan: 28 reference images. Okay, so it's got studio identity locks, tight headshots, profiles, medium, full body, neutral backdrop, soft cinnamon light. This is the spine. Real world context: truck cab, rooftop, kitchen table, front door, ladder, drive-through. Different environments, same Marcus. Expression study: tired but warm, determined, quiet relief. Focused thought, worn out, calm. Soul ID needs the emotional range to map to Vito's storyboard beats. Lighting variations: golden hour, overcast, harsh midday, low-key interior, cold pre-dawn, warm phone screen glow. The lighting arc of the entire film costs 6 credits each times 28, 168 credits. We have 2,994. Time: I'll submit in 6, in waves of 6 in parallel. So it is spun out 6 parallel agents. Hey, hey, hey, there we go. We got Marcus's modeling shots. All right, we are moving and shaking, guys. Who's digging this? This is awesome. All right, so we've got A1. So see how it creates a naming convention? So we understand we have character A1, we have a character named Marcus, and we've got process A, B, and C, and D. And so we are generating all of these different expressions. So we know now what Marcus looks like. We're getting different environments, We're getting all these kinds of different kind of scenes. And so this is all about Marcus. The fact— if this truck cab doesn't look great, that doesn't matter. What matters is that we're capturing and we're doing an iterative process. So yes, no, Marcus is probably not going to climb a ladder like this, but it's not focused on that. It's focused on trying to get consistency in our character Marcus, right? And so we're just trying to get a lot of consistency there. Okay, so Marcus fully cast, 28 references, 0 failures, $1.70 in credit, 6 minutes wall clock. The contact sheet for the audience. All right, so this is what you guys are here. So it's based what the audience just learned. This is professional AI casting looks like. Not type prompt, get character. It's a deliberate training set, same identity, varied conditions, designed to teach the model who Marcus is so it can be placed— so he can kind of be placed anywhere. So now I gotta approve the cast. So I'm gonna say, okay, we're gonna approve the cast. You want to generate any specific images? No, I think we're gonna go good. So we're gonna go 1. Approve the cast. Great job. Always say thank you because this thing's going to have a body soon. All right, who's digging this? You guys enjoying this? Is this helping you guys? All right, so we've done— now here we're going to upload the 28 images to Higgsfield Soul ID. 10 minutes train time free with the team plan. So Soul ID, where is that image? Higgsfield Soul 2.0, designed for creative fashion-forward results, now available. Soul Hex, again, I don't know what the heck this is. Oh, this is Soul, designed for creative— what is that? Assets, Cinema Studio, Mark Character, Soul ID Character, Create a Character. Okay, so I don't know if we can do this, if it can go Higgsfield AI character upload. So we're going to ask it. Pass approved, onward to Soul ID training. Let me first verify what character training endpoints the MCP actually exposes. The The Soul ID training might be web UI only, or it might be wired through media upload. Okay, so I might have to use the UI for something. That's what it's telling me. I might have to use the user interface for something, but I can do that. It doesn't matter. Hey, we're going to put Matt's stuff up on the screen here. Matt, there you go. Tap out. Okay, so important finding to be transparent to the audience about: Hayfield MCP doesn't expose Soul ID training directly. The formal training character step It's currently a web UI only action. So we have two pro-grade paths. Path A, the cleanest: upload all 28 images into Higgsfield's workspace via media upload, and then he clicks train SoulID. Okay, path B, the working pro alternative, what most production teams actually use: multi-image inference time onboarding. Pick the 4 to 6 strongest anchors. The long— for the live demo, I'll do both. Upload to the workspace. Adam can train SoulID later if he wants and select the top 6 anchors for tonight's shot. We're gonna do it the best way. So it uploaded them all to where? Upload all to the workspace into the Adam's Higgsfield workspace via media upload. Then he clicks train Soul ID. Okay, so we're gonna go back over to Higgsfield and we're gonna do this together. So we're gonna go here, present. Share screen. So we have it in assets— maybe not yet. All right, my folder, all assets. You said you uploaded those, all those individual images into the— where is it here— into the workspace via media upload. It's a workspace. Where is workspace on here? Nope. And open up a new tab. Where is here? Canvas, Explore, Workspace, Workspace. See, that's why I hate using software. So you're gonna learn so many new softwares. Not in here, not in here. What is this? Photo dump. What's that? Generate any aesthetic? No. Edit workspace canvas. Audio collab video. Again, I'm figuring this out with you guys. Image. Is there something in here says workspace? You guys see it doing the upload now? Okay, so it's saying it's doing the upload now. All right, let's go back to the chat. So I'm just watching my, uh, Telegram over here. Ruf Club Claudio here is uploading it to the media library, says it should take about 30 seconds. And so that's why I want you guys to help me out in the chat and keep me, keep me going here, because I would love to see what you guys, what you guys are interested in, what you're learning. What is the video that you would like to create? Have you ever watched an ad where you were just like, man, I would love to make that ad, or we need that ad. What do we got? What do we got? You have 3 seconds. Let's hop over to assets again. I think that's where it uploads them, is into assets, but it hasn't done it yet. What's this do? Ah, cool, makes it bigger. This is— okay, so all right, I'm going to show you guys what I got back from Roofclaw here. This is the coolest thing about it, is like it just— it makes everything so much easier. Life is easier with a Roofclaw, guys. That's just what I, what I can tell you. It's worth the money. All right, so for the audience, the Pro pipeline is two-stage. First, you generate the training set, then you upload them to your character training system. Not just a series of one-off prompts, this is how agencies build brand films to scale. Once the cast is locked, you can produce dozens of scenes in a day without identity shift. So roll into step 3 now. So what do we got here? So he uploaded 20 to 28. Look at Higgsfield. You see them in your media library. Go to Soul ID character training, select all 8 images, 28 images. Okay, so let's, let's, uh, let's see if that's the case. So there's green. Hop in Higgsfield. So is it in here? It's not in here. So we go to— oh, there we go, we're in here. So we have all of our assets in here. Now we go to SoulID character, right? That's the— SoulID training is now possible. SoulID character training. All right, create character, upload. Okay, but I don't want to do it from upload. How do we do it? Upload, avoid duplicates. Maybe I do it from assets. Oh, I got download and upload. That's dumb. That's stupid. All right, let's make this small. All right, Mark, so we still got to do some clicking around. This is the process, right? We gotta— we're gonna— once we have our character though, because I'll be able to cast Marcus again anytime I want I just have to go through the process of downloading him first, right? I'm thinking for Mother's Day I should send my wife a video card from Marcus. He's just way more handsome than me. All right, so we're gonna go download. It's gonna zip it all up. We're gonna unzip it. We're gonna put it in, and then we're gonna upload a Soul ID. Now, the MCP for Higgsfield, like, literally just dropped yesterday. So why is this happening? There's kind of a cool story as to like why this is happening. So Facebook bought a company called Manus, which is like a super agent, and that was a Chinese company, and they bought it, and then what they did is because the United States doesn't really want like Facebook, a giant gajillion dollar company, to have all of their stuff sitting in a Chinese, like, headquartered company. So they moved it to Singapore, and then all the Manus employees moved to Singapore. And they treat that like— and China treats that like Canada does. Like, Canada doesn't love it when, um, all of our smartest people move to the United States. That's like the thing that they complain about. And so what they— so they don't like that, so they call it like a brain drain. So they had brain drain to Singapore, so China canceled the deal. $60 billion deal. They canceled the deal and said, hey Facebook, actually we're gonna unwind that deal, you can't buy Mannis. So Facebook said, well fine, then we're gonna make our advertising API available to everybody with all kinds of things, which has just happened today. And so now anybody with any agent can access the Meta Ads API, or you could just write your own workflows in N8N or something like that. And so we're gonna go and say Marcus. And then we're going to use Soul 2.0 or Soul Cinema. What does it want us to do? So go to Soul, study character training. Okay, so I'm going to stay here. Before I go forward, should I use Soul 2.0 or Soul Cinema? Which one is the newer version? You can probably tell by picking the credits. If I go Soul Cinema, is it more credits? No, same credits. I go Soul 1.30 credits. SOL 2.0 is more. So I'm getting Claudio to do some research for me to find if I should use SOL 2.0 or SOL Cinema. Now it says it could use up to 80 images for training Marcus, up to 80. So I'm still not doing it the best that I could be doing it. But again, we're trying to take you guys along for the journey. So hopefully we're gonna get a good result. But good is all I got for 28 out of this. So we're gonna see what we do here. I'm just waiting for, uh, waiting for Claudio to tell me how it's going. So for those of you in the room, if you just joined, thank you for joining. Do the like and subscribe thing, it helps me out. It makes me want to do lives at a time other than 9 o'clock at night Pacific Standard Time, because if you guys dig this stuff, then I will do it more early during the day. But this is me working through something I want to do If you just joined, I'm working on making a Super Bowl ad using Roof Claw, and we are going to turn it into a video, and then we are going to put it live on Facebook as an ad. And we're essentially going to go through the, the video ad creation process to actually just automate marketing. The first time is the hardest time. I've never done this before. I'm just doing this based on my experience with AI and walking you guys through the journey. So, all right, so, so bonus. So what did it sound here? So it said, to honest, Soul 2.0 is the newer numbered version, but they're not competing versions of same thing. They're sibling models with different jobs. Here's what Higgsfield Catalog says: Higgsfield Soul 2.0 is realistic UGC, fashion, editorial, and character generation. Soul Cinematic is cinema-grade stills and concept art. The version question: Soul 2.0 explicitly carries the V2 tag. Soul Cinema doesn't carry a version number. It's different product specialization, not an older Soul. Think of it like Final Cut Pro versus DaVinci Resolve. My recommendation is for the Ode to the Roofer demo. The pro move: use both picked by shot type. All right, bonus catch. I noticed Cinemax Studio 2.5 is also in the catalog. That was the model we used for V1, V2 in our last session. It's specifically tuned for cinematic stills up to 4K. It's actually a stronger pick than Soul Cinema for our hero. So use Cinemax Studio 2.5 at 2K. Well, I don't have that one, so I'm assuming that that's something that's coming soon. So given that Soul 2.0 is at the top, we're going to do that and we're going to follow its advice, and we're also going to do that as well. We're gonna, we're gonna try both. Why not? Let's burn the credits, right? Our self-worth is measured in the tokens that we use every day. So we're gonna go like this and we're gonna go Marcus Cinema. And that way we can see the difference. Right? I crawl so you, you guys can run. Alright, so we're going to create Marcus and Marcus Cinema. Alright, so now we're going to go back up to our plan. So step 1, step 2, we've cast Marcus. Step 3, generate hero frame, the opening image, Cinematic Studio 252K. Prompt locks in the 5 layers from the doc. White pickup, dump trailer, emotion. Okay, so, all right, so both of the, um, both of the models are training right now. So I have, um, I have Marcus, which is trained on SOL 2.0, and then I have Marcus-Cinema, which is trained on Soul Cinematic. There was no Soul Cinematic 2.5, or at least that's not how it was versioned or displayed in the user interface. I think that Cinematic Studio 2.5 is probably for generating images and not making SoulID characters. We'll figure it out. I got two versions, we just gotta wait for it to train it. They're both in queue. So walk me through for the audience what we're gonna do at step 3 for the hero frame 1, and make sure if we need Marcus to be done training that we can provide the audience with some value in the meantime. All right, so coming back to you guys, um, so I just sent that over. All right, we, you know, I got you guys here in the chat. Let me know what's going on. What are you guys thinking? What are you guys seeing? What do you guys want to— what do you guys want to know? Um, all right, we are 2% complete with Marcus and he is in queue. So we are training this. I'm going to switch back over to Claudio. I wonder, is there, is there a way to— no, there's only— can I do extra camera? Probably not. That's like an extra camera. It'd be cool if I could have two screens loaded up sharing so I didn't have to kind of bounce between both of these. All right, so smart captions cinematic 2.5 thing, blah blah blah, even better. Step 3 doesn't actually require Marcus. Panel 1 is an extra— is a wide environmental establisher, no face in frame. We can execute it now in front of the audience while both Soleidis train in the background. That's the move. So what we're making, all right, so what are we doing next, right? We have— we're trying to make a Super Bowl ad, right? So we got to follow the Super Bowl ad framework because if you guys are making videos for roof— for your roofing company, you got a couple choices. You can do a drone shot of a video, which is mark, mark shitty, or you can generate Facebook Lives, which is your best. You can make another reel with pop, pop, pop, pop, pop, teleprompter along the bottom and just trying to do your retention editing where you zoom in and you zoom out and you zoom in and zoom out, right? Or you can make something cinematic. You can make something that tells a story. If you can make people laugh or cry, you can, you can convey something to an audience. So that's the whole idea here. So the first scene we're doing is, is a 6-second scene. Drone descends towards a single white pickup and dump trailer on an empty strip mall parking lot. Pre-dawn, 5:47 AM, fog on the pavement. The interior cab light is on. The whole world is asleep except for whoever's in that truck. So that's the scene. That's not the image generation prompt. Why this is the most important frame in the entire film: this is the emotional contract with the viewer. Before they hear a word, before Marcus has a name, they recognize themselves in that lonely truck. This is you guys, right? If this frame fails, nothing else matters. V.O. Storyboard literally calls this out. We identify with whoever is in that truck before you've ever even seen their face. Right? This tells a roofer, this is for you. So the directional translation, storyboard, 5-layer prompt. Okay, so it's giving— so this is the idea, the 5-layer prompt. So these tables don't format nicely on here because you can't like slide this over, but subject, anchor, white pickup truck, it's basically creating a table and it's going to create this prompt for us. And panel 1 is cooking, 25 to 45 seconds. So what I want to do is I want to share with you guys— I want to share with you guys a video Okay, that demonstrates my point of what you guys want. So we're going to go to— all right, we're going to show you guys the video, um, and you guys going to see what I mean when I say You got to make people laugh and cry. All right, so we're gonna stop screen. You guys got 2 minutes? Of course you got 2 minutes, you're hanging out with me. Also share tab audio. All right, we're gonna do this. Zinn and Lock In, you're gonna cry, boys. By the way, what do you see down here? HubSpot chatbot, baby. Alright, so imagine you as a roofing company, and you set that, okay, you send that when you— as a thank you video when somebody signs a roofing job, and you send it again after they're done, right after they get back. Tell me, tell me that you have not hooked them emotionally. Tell me you have not made yourself memorable. For the record, I know that roofing company and they make hay. So they do. And that's the point, is that you want to make your customer fall in love with you. All right, you want to make your customer fall in love with you. And so I want to make roofers fall in love with me by feeling understood. And so That's why we're making the ode to the roofer. So what we're here now is like, while we're generating this frame, right, two character models are training Marcus's identity. Here's what's happening under the hood, because this is the magic that lets Hollywood-grade AI ad production actually scale. SoulID is taking the 28 reference images that we just curated and learning what makes Marcus Marcus— not just his face, but his facial geometry, his build, his wardrobe, his family, his energy, The model's essentially building a mathematical signature of the identity that survives any environment, lighting condition, or camera angle that we throw at it. And why this matters is because without this step, every shot in the film generates a slightly different person. The audience never bonds with a character. They bond with a consistent face. SoulID is the bridge between AI looks cool and AI made me cry. So we're training these two versions because they specialize differently. Marcus on Soul 2.0 is realistic user-generated content. Best for the tight intimate portraits. Markus on Soul Cinematic is cinema-grade dramatic atmosphere, which is better for the wider lit-up emotional beats. And so now we've developed the keyframe 1, and so the whole idea is, as a roofer, if you're a roofer in the room, could this frame with the sound off telegraph the emotional state of the entire opening of this film? Do you know that this is about a roofing company, or possibly whatever company, from looking at it? We got the single cab light, We've got the truck, we've got the dump trailer. Now, was this what we were looking for, right? Like, is he parked funny? Yes, right? So we have to look and say, okay, is that what we're going for? So pre-dawn, Marcus in a truck, right? Phone will go on face. We said we didn't have Marcus yet, right? So now we don't actually need Marcus. So what we can say is like, well, Marcus isn't in the truck, right? Marcus isn't in the truck. So we could say Hey, uh, the one thing I noticed is that Marcus isn't in the truck, probably because he hasn't done his training set yet. But I want to know, do you, right, need him to be in the truck? Should you generate more versions, can you do that based on a description that you have of him? But the fog is there. It's not a wall of fog, it's just a light fog. So everything you're saying— and this truck does feel small in a vast lot, you definitely got that part right— but the truck does dominate the frame. So it's awesome. The only thing you're missing is Marcus in the truck. I don't know if that's important to you, but maybe you generate 2 more versions before we ship it. Okay, so that's what we're doing, that's what we're doing right now. So I want to go back up to our plan and see what our steps are. Okay, so here's the plan ready for stream. So we generate the hero frame, right? Generate hero frame 2, the final, the final image. So this is the final image, same prompt structure, same composition, same truck, Same angle, Marcus visible inside phone screen lighting his face, emotion. Okay, so we can't do this yet because we don't yet have Marcus's character generated. Because if I go here and stop the screen share, let's see where we're at with our training. So again, once Marcus is trained, you can use Marcus in all kinds of videos. That's the magic, is that you start to create episodic videos. All right. Oh, look at this, we have him trained. Both are trained. So we actually have Marcus at this point. So now if I go generate, what happens? Okay, so we have him here. So now we have something we can do. All right, all right, because this is back in our assets, right? Yeah. Is this— you guys digging this? That Jack thinks is pretty cool. That's awesome, Jack. I'm glad you're still here. I'm glad you're still watching. All right, so we're going to go back to— we're going to go back to Roof Claw and see what's going on. All right, so share screen back over here. What do we got? All right, honestly trying to know— Vito's storyboard actually calls for no driver visible in panel 1. The artistic logic is to identify the idea of someone in the truck before we attach a face. A specific Marcus equals audience match. It watches Marcus, anonymous silhouette. Audience is market, but you're right, and that current frame is missing something. The cab light has no source. It looks like a parked truck with a dome light left on, not an occupied truck. The fix isn't necessarily make Marcus visible, it's make the cab feel occupied. All right, so my pitch: silhouette, match the storyboard intent. Version 3: hinted Marcus, your direction, same drone wide, solid consideration Marcus isn't done training yet. Well, he's done training now, bro. Damian, it's— this is really cool. I'm telling you, this is going to be the best. This is gonna be the best YouTube and Facebook live stream for roofers who want to use AI. I might not be Alex Finn, I might not be, uh, I might not be Matt Wolf, I might not be Matthew Berman, but I will be the best AI channel for roofers because this stuff— I can stay up till 10:07 PM at night standing in my home office, you know, just, just cranking on this kind of stuff with you guys. And I hope you guys really enjoy it, and I hope you guys learned something. I hope you guys feel empowered to go take action in your businesses. And so just as a reminder again, guys, if you guys are watching this and you guys want this skill, I am putting everything— so every time I do these streams and I build something cool, I put it up here. So if you want the Roof Claw Local Service Ad Agent, a service, basically a voice agent that is going to pull Google Local Service Ads and then go contact your customers by phone and text message, you can just point your Roof Claw at this skill. That's the whole point of Roof Claws. It's not just another open cloud. It's an open cloud trained to roofing, trained to your business that you are trained to use. And then the other part is that it, it comes with expansion packs, right? And so we are, we are turning everything that we are doing today into another GitHub repo that is public, that is free to use. There's no gate behind it. We're actually going to blog about it when we're done. But this here is the Roof Claws Meta Ads agent, and it's going to come with the Higgs field AI training thing with how to make Super Bowl level videos. And that's why we're putting through these pull requests. And so we are going to update the code base here in just a moment, probably, because if I look at our status— oh sure, sorry, share this tab instead. This is the Roof Club Meta Ads agent, and this is going to come with everything. And so you can see Claudio's in here, I'm in here, Vito's in here, and we're all updating this as we go. And then you will be able to point your agent at this and run through the Launchpad Markdown file which will literally guide your agent through building this with you. And so that way you will have this added to your system. That's your expansion pack. All right, and so if I go back over here to assets, let's see, did we get some more? Not yet. He's, uh, he's creating some more right now actually. So we're going to jump over here. Yeah, or security consulting. Yeah, UK, if you're— oh Oh, you lost connection to my screen share audio. That's okay. All right, so let's jump back over to Claudio and see where he's at. All right, so, all right, so, okay, so it's got me some more renderings. All right, both landed downloading and viewing. So check this out. Okay, so it's hinting at Marcus, slightly older truck. Okay, different truck. So we're gonna have to make sure we get some truck consistency, right? All right, so fog is seen, setting drifted, looks more like a misty forest road than an empty parking lot. Yeah, so like this image is better. So if I had to say like which is Marcus, it's probably this one, right? Neither is shippable yet. V2 has the right intent but wrong setting. V3 has character but loses the poetry over restraint of the storyboards because it's going to do version 4.0. Microinnovation V4, the synthesis. Roll V4. One thing I think we need to do is get some truck consistency, because in the version 2 and version 3, one has a super old, like, '80s style Ford truck and then you have like a '90s style Chevy truck, but in the original you had more of like a 2014 Ford F-150 or F-250 single cab long box truck. So we need to get some truck consistency built into our video here so we don't confuse the audience with multiple different trucks. Um, I haven't seen a version 4 yet. I only see the 2 versions with the Chevy truck and the 2— you sent me 2 images with the old Ford and then you sent the, the original photo, which I liked the best. It was the most cinematic. I would honestly use the very first image if we had no others. The only thing it's missing is a dude in the truck. Let's see. All assets, nothing yet. What do we got over here? Okay, so when we see, we're going to see version 4 maybe, um, coming through, coming through maybe. All right, and so you're right, got those V2, V3 trucks drifted to different vintages. The fix is to lock the truck spec explicitly and use V1 itself as a style composition reference. V4, use Cinema— V1 cinematography. V1 is a reference image. Explicit truck spec, faint silhouette. 2014 white Ford F-250 single cab long box, 2 doors, primary factory chrome grille, factory steel wheels, lived in working condition, attached to a matte black single axle dump trailer with a folded blue tarp strapped on. Probably would have had a double axle if I could pick my dump trailers, mate. But we are probably trying to paint Marcus as a sad story. Startup roofer switching. Nano Banana 2. All right, so we are— so basically, see this? Like, it's working through the problem, right? It worked through the problem for me and says, okay, I'm actually going to switch to Nanabanana 2 to get the better output. And so it's doing all the figuring outing for me. I'm not sitting here banging through, banging through, um, uh, Higgsfield the whole time. Like, yeah, I gotta hop in there a couple of times. I get to go check out some stuff, but it's not sitting there making me suffer through it. All right, all right. So I see a few more people that popped in. Where are you guys coming in from? What are you guys looking for? We are generating a Super Bowl ad and running the Facebook Meta ads. So we are going to generate a Super Bowl ad, we're going to post it to Facebook, we're going to get as far as we can. I want you guys to enjoy it and get some value out of this video. So I want you guys to tell me where you're from, what's your roofing company, where are you located, what's your struggles, right? We don't care about your struggles today, we'll leave that for a Winning Wednesday webinar. Well, let's see, we got Nano Banana cranking out. We're burning some credits. Northern Wisconsin, Roach Roof, nice to have you here, man. Thank you for being— is Roach— yeah, it's Roach Roof, Roach Roofing Company. So yeah, I, um, I loved it. You guys in Wisconsin, so you guys got hit by a storm lately. So what we're doing right now is we're working through a storyboard process. So what you're seeing here is Roof Claw Roof Claw is essentially a custom-built version of an open-source software called Open Claw that is customized to roofers, built to support roofers in their business. So it's like an employee, it basically sits on a laptop. If you come in behind me here, you'll see I'm surrounded by MacBooks, and that's because I build these and ship them to you. And so I have all these MacBooks around that I'm using, um, and it's basically a little bot farm. And then I send you guys the cattle, and then you guys get to use it. All right, so we got Marcus in in a truck, and it's the same truck. So this is good. We actually got kind of two images, similar truck, similar to— so we're good. All right, all right. So both of those shots actually work really well. I'm gonna save them just so I have them on my computer as well. Oh yeah, buddy, we are devouring some tokens today. That's the whole goal, right? Your self-worth is— your self-worth is now measured in the amount of tokens that you use. So I'm going to Save As, and we're gonna go Marcus Close Shot Truck. So I'm gonna save these because I kind of want to, like, I kind of want to be able to use them again if I have something else. Maybe save them somewhere else, you never know, you never know. There's our Marcus V1, Marcus Truck Distant Shot. All right, okay, so we've got our opening shot. What's the next step? Um, we can scroll up here back to our plans. We had Marcus generate the casting character. And then where are we at? We're going up here. No, whoop whoop. All right, so generate the hero frame 2, the final image, the mirror, same prompt structure, same composition, inverted emotion, same truck, same angle. All right, so it looks like this is where we're at. So we're on step 4, which is the final image. So say, sounds like the next step is step 4, which is the final image. And so I guess we're going to go back and fill in the in-between, but this is the final image, the final shot, which completes the story loop. And so I want you to finish up with step 4. Here's a reminder. All right, so I'm going to go back to— we're going to go back to, uh, to Higgsfield. And we're going to go here. I really got to figure out if there's a way to swap between screens. All right, so if I go back to assets— so it's not in here. So I'm— I think it actually generated with Nanobanana through my other API key, but that's okay because we can upload— we can upload assets. So if I go— how do I go assets? And I can go Cinema Studio. Oh no, they're right here. Okay, Marcus is sitting in the— whoop, this is a bad one. So how do we How do we say this sucks? Delete. And we only want the Ford images, so we're going to get rid of this one. We're going to get rid of this one because this doesn't look like— this doesn't look like the Ford. We're going to get rid of this one. Okay, so we've got our cinematic shots. And oh, I didn't send that message in. So what is the video that you would like to create? If you kind of— if you had a video idea, if you could describe a video in 3 sentences that would make a customer— make a customer cry or laugh, what would it be? Because I've seen some funny ones too. You guys want to see some funny AI-generated videos? I'll show you some cool ones. I got one here. We've worked with a few customers on this in the past before we had an agent to do the work. Uh, let's go Sweaty Italian. So we got, we got Roof Claw busy working away. He's typing up here. And so this is actually communicating with the Mac Mini on my desk if you guys weren't there. All right, we're gonna switch over to share this screen. And give you guys a quick laugh while we're waiting. I gotta entertain the people. We gotta entertain the people while they're waiting for AI to do its job. So we'll show you some stuff we've done before. All right, so make sure we're sharing. I gotta make sure I'm sharing the audio. Actually, share screen and share the audio. Yep. Okay, so while you're waiting, we're gonna go and watch this Sweaty Italians. Here, all of this could have been prevented. All they had to do was spend $199 to clean the damn dryer vent. Pause. You're probably wondering how it got to this point. Let's go back a few days and see where a family made a stupid mistake. I was knock knock knocking on doors, helping families prevent the number one cause of house fires by cleaning out the dryer vents for only $199. This family said no. And look what happened. I regret to inform you the family didn't even die. For the record, we've collaborated with this guy, $2,500 to make this video even started. To think all this could be prevented. Big words. $7. Click the link. And we stopped by to clean the dryer vents without them even knowing we're there. $199, no carbon monoxide, no house fires, no face-to-face contact with sweaty Italians. Sounds like a deal. Click the link below. Let's avoid all the drama. So this is what I'm talking about, guys, right? We got to make customers laugh and we got to make them cry. And so you— this is— that was just a storyboard, right? Like, that's 100% AI-generated. AI is like plastic surgery, you only notice it when it's bad. But that was funny enough that people will forgive If you're funny, people will forgive. So you got to make it funny. And the great thing is, is AI is a great ideation tool. And when you have a Roof Claw and you point at this GitHub repo, it will guide you through the story, this, this, the SumoQuote ad ideation process to come up with something funny. So even if you are not funny, you can come up with something cool. All right, so we got— we're cooking again. Claudio's busy. Let's show you guys what Claudio is up to. All right, so we've got understood wide to wide. So this— so we're working on the next things. We got wide to wide mirrors the graph shot, the graph— get the gas move, same lot, same truck, same drone heights as V1, but the world is flipped. Cold blue, warm golden slump silhouette upright with phone screen glow on his face, the fog thinning as the sun breaks. For character identity, I'll multi-image condition Nano Banana 2 with 4 anchors chosen surgically for what each contributes. All right, so V1 itself, composition. C3, quiet relief. D6, green foam glow. Okay, A1, front headshot, pure identity lock. So Marcus reads as Marcus, pulling the URLs. You know what, I think I haven't told it that Marcus finished. So, all right, so while it cooks, so what you guys need to know, this is what you guys did. This is the important, most important shot in the film because the visual rhythm, the visual rhyme, it's the same lot, the same truck, and the same drone height. The same camera angle, but the world has flipped. Remember what I said earlier about the Save the Cat story beat sheets, right? It's like, uh, you are the hero. So in my video, you the roofer are the hero, right? And the hero has a problem, right? There's a conflict, and it is either internal, external, or a philosophical problem, right? And then there along comes the guide, right, which is me, RBP Roof Claw. Right? And the guide has a plan and a promise. And the promise is to solve the problem, um, to remove the pain— internal, external, philosophical. And so then there will be fun and games and a climactic event where the, where the problem is now gone. And the— either the world has changed around the roofer, or the roofer has changed in the world. That's the whole point. And so the whole idea is that we are changing it. So when this lands and we cut between V1 and this frame, the audience gasps. This This is not me trying to be dramatic. This is exactly how Whiplash and Manchester by Sea and The Bear all build their final beats. The same composition repeated with one variable change creates an emotional voltage that no other shot can deliver. So notice what we did with the multi-image conditioning. We fed the model 4 reference images, each contributing something specific. Our V1 frame— this is for composition, the lot and the truck identity. 2, the Marcus headshot A1, who he is. Marcus C3 is the quiet relief. This is his expression in this moment. And Marcus D6 is the phone glow, which is the exact lighting reference. This is the layered prompting that pros use. So you're not throwing one thing at the model and praying. You're getting— giving it a reference or a director's reference packet. And so it doesn't— so it doesn't like what it created yet, doesn't gasp yet, right? So he's smiling. There we go. We fix the gasp with two surgical changes: lock composition, emphasize identical drone altitude, force the phone underglow. So I want to say here, try V3 with Soul Cinematic if your Marcus industry— so I gotta let it know. Yes, the Soul Cinematic has finished. We now have that character in Soul Cinematic now. Ah, crap, it's not school cinematics, soul. By the way, this is your— by the way, soul cinematic. So basically letting it know that we do in fact have the soul cinematic is done now. A little bit morbid I think to resonate. I don't know about you but— I know what you guys— but I go through washing machines and dryers like they're going out of style. You told me I would make my machines last longer. You told me it would make my machines last longer? I understand. A little bit more but to think everybody— but I, I think too— a little bit morbid I think to resonate. I don't know what you guys— but I go through washing machines and dryers like they're going out of style. You told me I would make my machines last longer. I don't understand what that means. Road Roofing, you're gonna have to— I have to— oh, I'd buy. I was talking about the $199 dry cleaning. Oh, right, right, right, right, right. Yes, 100%. Yes. So that's just saying you'd buy if we got the— if the dry cleaning made the machine last longer. See, so that's— that is a great example of like how you can create value in excess of price, right? If you make the washing and dryer machine— if you make the dryer last longer because you clean the dryer vent you make the $2,000 dryer last longer because of $199 dryer vent cleaning. This is value in excess of price. That's exactly what you're going for when you're doing something like this. All right, so what you're seeing here is you're seeing a bunch of code and gobbledygook, right? And so it is going to try and say, all right, I'm going to try and use— you do these surgical changes. So I'm going to copy because I think it's using Nanobanana to generate images rather than actually using our Soul cinematic, but that's because I didn't tell it. Again, because it's all about prompting, you got to tell it what you're doing. And so we have our Soul, so hopefully we can use Marcus with Soul 2.0 to generate. So I don't know how you— like, how it pulls that in in the MCP, right? So maybe we could search that, right? So I'll come over here and we'll do Soul cinematic. And we're going to share this tab instead. We're going to say, can you expose the endpoint in the MCP for Higgsfield that allows you to put your soul characters into your videos as opposed to just raw images? Because again, this MCP like literally got launched yesterday, so it's only going to get better from here. They might not have exposed every single endpoint or every single tool, and so now we might have to actually, you know, figure this kind of stuff up. Okay, so prompt. Okay, so here we go. So custom reference ID, your character ID. So it has to know which character ID it is to know which one this is. So when you train your character in Higgsfield, often referred to as a Soul ID character, the system generates a unique identifier, right? Good data model. Okay, so we just have to know what is our custom Soul ID. So if I go in here and go back to character, Soul ID character, I wonder if I go inspect Go a little bit like this. Why can't I move this over? There we go. Oh, I go back, back. What I'm doing is I'm just looking for— seeing if I can find the soul ID on here. It's probably going to know it. I'm just like, I'm just interested in how Higgs field works because then I'll know more about it, but I don't think it's actually exposed closing it here for me. We're gonna go to class overflow header promotion style style main. I could probably search someone smarter than me. I got a guy works for me named, uh, named, uh, his name is, uh, Oh, come on now, Adam— Jose, he's even smarter with Chrome DevTools than I am. Oh, wait a minute, what are these three dots down here? Edit cover, reuse, update for cinema. Edit cover. No, let's hope I can get cover character uploaded. No. Knows Marcus. Okay, so firing— okay, so it knows it's firing Soul Cinema. Okay, so it's got it. Sorry, I'm not sharing that with you guys. All right, we're gonna go here, stop share. Go back to our Roof Claw. Excellent. Soul Cinematic and Train Marcus is the dream tool for this dramatic lighting frame. Let me check, check exactly the Train Character Invoke, firing Soulstone D6, Soul Cinema. So I don't know, is that good Marcus? Not a great Marcus. I would try that one again and make sure you use maybe the other Sole 2.0 version of Marcus rather than, uh, Sole Cinema. Are you able to echo back the Sole IDs that you're using? Because I can't find the Sole IDs when I'm looking at it through the user interface, but maybe you can tell me the Sole IDs so I know that you are in fact using one or the other. We could also be running up against some context here at this point. Or you /status to find out like how many tokens have we burned. 183,000 tokens. We're not even at our million token context limit yet. We've got— so we've got— we're caching 99%, so this is good. We're running on Opus 4.7. Gosh, I gotta update my OpenClaw. Okay, and then we have Okay, so the truth, they have not been using your trained SoulID characters. I've been passing reference image specifically. Okay, so no, we got to be using as values here. Okay, so the train Marcus on Soul 2.0 and cinema on the web UI never been called. The reason the portraits look like Marcus so far is the reference images themselves are Marcus, so the model is faithfully copying. Let me investigate the actual mechanism. See, so there we go. Because what's this? Because we just call from a phone. Are you actually talking into a microphone and prompting Rufus? Are you just talking out loud what you're typing? Ah, sorry about that. Here, let me make it— let me make it bigger. I— there's another way for me to do this, which is like this or this. No, maybe this. This— which one is the biggest? That is probably the biggest I can get it. If I go like that, it's probably the biggest. Maybe if I went like this and went like this, that make it bigger for you guys? No, it doesn't. What if I go like this in Telegram? Can I zoom? Auto fill window, zoom. Nope, trying to see if I can make this bigger for you guys. Whoops. See, I told you, told you guys. How do I get this out of full screen mode? No. That's about as big as I can make it, my man. So yes, I did— when I dictate, you're going to hear a little click, right? Um, so if I go like this, you'll see her. Dope. Tuck, tuck, tuck, tuck, tuck. That's what you guys are seeing. Um, but yes, I am dictating to my roof claw. Actually, if you guys don't know what that is, this is a good time for me to get my $10. It's a good time for me to get my $10. If you don't know what WhisperFlow is, you're a sucker that's still typing like a loser. So I put my little referral link in here. So we're going to go here. And so if you check in the comments— that's on YouTube. How do I, uh, get it over to you on Facebook? So yes, I'm using something called, called whisperflow.ai. And if you use my, uh, I'll send it to you on Facebook because if you use my link, I get $10. Nothing like getting paid $10. Better than kicking the ass. I spent more than that in tokens tonight trying to show you guys stuff, so I'm gonna get that. Where are you at, Damian? There you are. Sent me an X post today. There we go. All right, at what point do you think this transitions from bouncing between models and endless tools to a Jarvis from Iron Man type assistant? So this is, this is already pretty Jarvis, Roach. Like, the whole idea is like you're— I don't have to jump between models. And like, I'm training this thing, I've never done this before. This will actually get turned into skills. So if you actually go into like— as I'm going like today, as I'm going, I'm actually pushing all of this over to this thing called GitHub. And so if you have a Roof Claw, this software that I can, I can give you, right? Or not give you, but sell you, you can buy it. If you get a Roof Claw you can point your Roof Claw at this system, okay? And you just point it to this link. It's all you're going to do. You're just going to point it at this link. Actually, it'll be this link right here. So you will point it at this URL, which I will post on the actual live video as a comment, okay? And so you will probably see it in the comments here. You'll be able to point your Roof Claw at this. It'll launch the launchpad file and it'll walk you through setting this up, and then it will be trained because I'm going to be putting all this in for you guys so that you guys don't have to go through all this, because I will put a lot of this practice— like all these little, these little blips that I've had, right, um, have been kind of breaking up. Okay, so here we go. So confirmed your Soul ID training happened. The Soul ID minus 25 credits. The MCP doesn't expose trained Soul IDs. Okay, so now here's what we're going to do. Open Higgsfield, click on Marcus, look at either the URL bar should contain a UUID, or B, click copy ID in the settings panel inside the character page. All right, so let's see if we can't— let's see if we can extract this. Let's show you guys some hacks to get around stuff. Okay, so we're going to do some hacks to get around stuff because I'm pretty sure if I come into Higgsfield, I can't actually get you guys. So see, remember I was using that inspect feature before? Because if I click on here, all right, if I click on here, it doesn't, it doesn't necessarily give me the Soul ID, does it? Nope. Main next static media. Command-F. So, okay. Go here next. Trying to see if I can find an ID. Soul ID consistent. We're gonna go, uh, too many IDs. So now instead what we're going to do. So I'm analyzing the, the HTML, but we're going to try something else. Okay, so I want to share you guys more of my screen here for a second so you guys can see what I'm doing. So we're going to go straight here, we're going to go share screen, and we're going to go entire screen. You're gonna get the infinity window, and we're gonna go over to here. Hey Claude, can you somehow find out the soul ID for either or both of the soul characters that I've created in Higgsfield? I need the soul ID unique ID files, but I can reference them in my designs. Because it's saying that if I give it the user, the unique, you, the unique ID, right, then it can use it, it can use it in my, in my video. So I, I'm using Claude inside Chrome. So this is the Chrome extension for Claude to see if it can dig around and find it for me and ultimately make my life a little bit easier. By getting me what I need. So that's what I'm trying to get it to do right now. It's not giving me the URL, right? I go over here. Yeah, it's not giving me that. Ah, see, it's using a JavaScript tool. It's gonna try and dig it out. And then there's another level of hack that I can do even on top of this, where we could use Codex to actually open this page. I'll actually show you guys that. Okay, so we're going to go use Codex. So let's rip this URL, and then we're going to go to Codex and say, hey, I want your help. No, we're gonna go over here like this, add a browser, I'm gonna go here, log in, continue with Google. And then it's going to ask me to open up my Smart Lock. There we go. Confirm it's me. Yes, it's me. Use the browser window. I want you to see if you can help me pull out of the JavaScript or the website code. I want to find the Soul ID for my two characters here, Marcus-Cinema and Marcus. So we're going to say, hey, can you rip into this web page and do it? Because Claude— there we go. All right, I have the soul IDs. So see, now I know how to find the soul IDs, boys. How do I go here? Stop screen share, go back to screen. So we managed to hack our way in and get it. So now back over here, share this. Got the IDs. Now we do this properly, saving them first, firing the genuine train character invocation. Ah, I say, let me— that this ain't stopping me. Now, so, but we'll jump over to Codex because it's also plugging away. Let's close this. So, all right, so, so what Codex is, by the way, what Codex is, um, if I go here while Claudio's working away here, um, go to Window, Codex. Where did it go? So here we go. Here, I don't know if you guys can see this window. I'm switched over on another screen, but basically it's inspecting the card containers and finding the ancestors. So it's basically hacking the site for me too, and it's finding what's underneath here. So Codex can actually like use your whole computer as well. So, um, it just— you can't fire it up, you can't talk to it from Telegram, You can't use it from your mobile phone. It doesn't have a soul, it doesn't have a memory. Um, it's a little bit different, but it's going to pull through. It would probably get it as well. And so coming over here, stop screen, go back over Claudio here. All right, so Solidity saved the project state. Now find the genuine soul invocation with soul ID Noice. All right, so what we're, what we're going to do is I'm actually going to come— once I get this back, what I'm going to do is I'm going to commit what we've learned to memory so that I save it on my machine. And then we're going to take the memories and we're going to commit a new pull request, a new submission to that GitHub page So then that GitHub page now has what we've learned. So all the crap that I went through today in the last hour and 43 minutes with you guys, you don't have to go through all this crap because it's going to learn, right? Because your Roof Claw actually has a browser relay. It has, it has the ability to do what I did with Claude, that thing that I did there. It has the ability to do that. Um, uh, It has the ability to actually kind of pull, do exactly what I did with Claude or Codex. It has the ability to do that. It would just, um, it would just— you wouldn't see it like I saw it because it would be happening on the machine over there. But I just did it so I could show you guys. All right, so what do we got here? So we got the wrong truck, so we gotta go here and say— oh crap, I didn't mean to do that. So, /plash, by the way. Oh, sorry, /plash, by the way. So I think you're doing a better job now. But the only thing is you lacked the truck consistency. So you have the sole consistency with a person, but you don't have the truck consistency in that most recently created image. So one thing I want you to make sure that you add to your skills when you are doing this in the future with me, and by extension when we upload this to GitHub, I want to make sure that you add a kind of a lesson, which is that you want to get consistency not just in the character but also in the scene, as well as with things like trucks and tools. Like, if we have a two-axle dump trailer, we always make sure that we get a JSON object so that when we're prompting Higgsfield, we get not just character consistency but scene composition and, uh, like props or items in the scene such as the truck and the dump trailer, or even things as simple as the phone, like if it's an iPhone or an Android phone. So that's what I'm doing is I'm going through this. This is the first time I've done this. I'm going through this and we're— and I am learning all this for you. And then I'm going to be uploading it to the GitHub project that you see there up on the screen. If you want a Roof Claw, you can scan the QR code on your screen or you can go to roofclaw.com and you guys can buy a Roof Claw on there. It is the cheapest, most powerful, smartest employee you're ever going to buy, and you only have to do one thing. Come here and watch what is going to be the worst orchestrated but most useful, uh, roofer live stream that you're gonna find on the internet. All right, see, so now we're building JSON. See, this is JSON. So the JSON says, okay, characters Marcus, vehicles Marcus truck, make model 2014 Ford F-250 Super Duty, configure single cab long bed, Wheels white, factory steel, no aftermarket. Grille factory chrome. Lived-in working truck, mud on wheel wells, tailgate stickers, light scuffs. Negative, no extended cab, no crew cab, no Chevy, no Dodge, no Hilux, no aftermarket lifts. Marcus trailer, single axle dump trailer. Color matte black. So it's basically building an array of objects so that it's consistent anytime it generates. Then every prompt gets assembled from the from the Bible, uh, not freehand written. So panel 1, panel 5, panel 7, panel 14, every truck shot pulls the exact same truck spec, no drift, right? Um, in the GitHub repo. Okay, so we're going to check out this. Exactly the kind of lesson separates a hobbyist workflow from brand film studio workflow. Logging it, it's now part of how I work. So we've trained it. All right, so I want to say this is a good place to stop and consolidate our lessons and go into the GitHub repo and make sure that we update it with skills and lessons so that anytime one of our customers points their Roof Claw at our GitHub project, it pulls in these lessons so that it brings these skills onto their machine so they don't have to teach it and train it the things that we taught and trained you. Oh, what's my voice profile today? There we go. All right, all right, what do we got here? So no fancy logos on the side of the truck. I can't tell you how many trucks in the city are using exact same, same template that I did on my trucks. Yes, when we got into the roofing game, everybody had the like little itty-bitty sticker on the side of the truck, and then we literally tilted our logo to the side so we could get it literally from the top to the bottom And, uh, now it's kind of like the standard in 2020— 2026. All right, so what it's doing here, see what it's doing, say implement, is building the documents that it is then going to push towards that GitHub project. So now you guys, when you buy Roof Claw, you can pull in all of these skills. The idea being that this thing is built specially for Roof Claw. This is built somewhere so that you guys have a— you get to take all the lessons from today and you guys get to absorb it into your machines. Now, if you have a Roof Claw that's a production manager, you are not going to give it the skill. You are going to buy another Roof Claw. So if you've already bought one, you get, uh, 25% off your second Roof Claw. So that way you get a deal to buy a second one so you can have your marketing Roof Claw. All right, so we're going to go back up here and check out our Higgs field just to see what's going on. All right, so we can close— we can close this now and get rid of that. Ungroup. All right, I'm gonna shut that down and come back over here with you guys. How many agents per roof clock? You can have, you can have a whole fleet of sub-agents. You can have 32 sub-agents. You just can't have them exist in too much conflict with one another. Um, there you can have lots of sub-agents that perform several, several tasks, but if you have a marketing one and an accounting one, they're going to exist in conflict with one another. It can be done. I, I have done it. I've had up to 32 sub-agents on one roof clock. I sold it to a billionaire, but it was a very complicated, very complicated memory system so that it had a way to, um, to basically put a concrete wall between those agents because they accessed a separate memory system. So that way they didn't absorb each other's personalities. So, but if you have an accountant agent and a, a marketing agent, the accounting agent wants to cut the marketing budget to save money and the marketing agent wants to spend all the profit, right? But like, and I, I don't know a lot of billionaires, Damien. It's, it's not, not that I know a lot of billionaires. I, it, like, he's not even a billionaire, but he's probably like close, you know, probably like half, right? But, um, but he, but I know him because of something else I'm working on. He's like, I paid him to be a mentor and advisor, and he thought it was kind of cool. And so he's like, will you make me one of those things? And so I did. And he's got like his own private server and all this kind of stuff. So his kids all have Roof Claws that can connect to their private server, and it's called the Ark. Um, he's kind of a, you know, wealthy prepper. Um, part of the— he's not part of the evil cabal that runs the world, but he knows them. All right, this is exciting. We're gonna see his pull request here in a moment. and we're going to update it, and then we're going to be able to crank away on improving. All right, see, so now there we go. So it's like, so what have we got? Session consolidation complete. What's now in production for any customer's roof clutter pull? New docs, Higgs field lessons learned, cardinal rule model to job matching matrix, 28 image character casting pipeline, MCP schema quirks. So you get all the quirks that I dealt with. You don't have that train, sole ID invocation pattern. Iteration discipline, storyboard to model transition, 6 anchor reference image creation rule pre-generation checklist. So Croft references added in README and the CDAN's playbook live at this gives a personal memory update so I don't relearn this in future sessions. What this means for our customers: when a customer's RoofCloud points at your GitHub repo and reads it as a part of their implementation, it now inherits the model selection muscle memory we earned the hard way tonight, saving them the same 50-credit lesson we just paid. The exact 28-page image casting protocol, all this stuff. So now I click on this and I'll switch over so you guys see what's going on here. So I'll go here and stop screen and share over to our GitHub. So now in here, right, you have now gained all this. So if I go up here into pull requests— oh, did it do— did you just update? Oh, I just updated it. I think it— so I'm going to go back and say, did you update it without creating a PR? Did that because I don't have rules set up on this GitHub yet because it's not a— I don't have, uh, merge rules set up because it's brand new. You're right, I direct push to main. That was wrong. Even though reversing now, revert commit on main. Push the same changes to a future branch, open a PR for your review. There we go. So see, I'm teaching it because it's, uh, it basically just committed to main. It didn't actually push a proper pull request. All that means, a pull request just means you can log the change, right? So you can log that change, and that way it just shows that it came in, um, shows that it kind of came in as a version. So we've technically done 2 PRs on this, on this GitHub already, um, and that this is just making it so that now we have one. So now if I click this and we come over here and share screen, we'll be able to review the changes. This, and this is just stuff like, it might not be something that you find super relevant to you, but version control is a huge part of like learning how to work with AI is like storing code somewhere and just telling it that you want to keep a good version. So here it creates a description, a summary of what it did, right? And so tone matches the rest of the docs, code JSON example. So this shows like, hey, what changes did we make? So files changed 3, and so these are all the files that it added. So this is the markdown that your agent is going to learn from. So your Roof Claw is smarter than what my Roof Claw was because it wrote 244 lines of code that yours doesn't have to. And that's just one, that's just one of them, right? Because we go here, let me go here, we've got this here. And then so it says, hey, this is your playbook, you can read this, you can go read lessons learned. So you can learn the lessons that we learned. And then it added, deleted these two lines and added these four lines into this document. And so this is the changes that are done. I can come over here and go ready to merge, merge this request. Now bingo, is now merged onto the main, and so you guys have access to it and we can come back and see. Oh yeah, it's merging right now, so it's going through the process. I go here, there, it's gone. We have two closed merges. That we ran today. And we're going to do some more. Alright, so we're going to come back over here, remove this. I'll go back to Claudio. But this is how we eventually get to the point where this is just a skill we run. Alright, so let's check our context status. All right, we're good. So we're going to scroll way up and we're going to go here and we're going to see what we should do next. So Cast Market, Hero Frame. So we should do Hero Frame 2. All right, so you are currently at 237,000 tokens of a million or at 24%. We haven't done any compaction. I would love to generate the next scene after the first scene according to the storyboard that Vito drew up. I wanna start generating the scenes. You could do this with subagents, but I think you kind of have to pass through what's generated from scene 2 to scene 3, from scene 3 to scene 4, from scene 4 to scene 5. So I don't think you can do this really with subagents. I don't think we can parallelize this. So we probably have to just do one after another. Maybe what you could do is generate the prompts for scene 2, scene 3, and scene 4 so we can review them, just so I can show the audience what this process looks like. By the way, I just got to say, guys, like, it is probably 10 o'clock —or is it probably like 11 o'clock, midnight, 1 AM, 2 AM for some of you—I appreciate you guys hanging with me and throwing some comments in the chat and participating. It's been fun working with you. I really want to make these lives into something that's actually tactical. You know, I could, I could do the thing where like the YouTuber goes in, clicks around, shows you what you could do, but doesn't actually show you like the real hard battles of going from like what you could do to how it actually works. And instead of making the super polished video that like just shows like, you know, this destroys everything, MCP, you know, this is absurd, right? And then you just click on the video and it's a guy like setting it up for the first time and he just kind of shows you what it could do and then doesn't actually show you the hard battles. That's the thing that I don't love about YouTube, right, is like I have to watch 10 YouTube videos to get a grasp on, on this, um, and then try and like turn that into something that I can use. And I thought, why not just take people along on the journey? And so, yeah, so this is just, this is just me being raw. So raw is best. I love it. We're raw doggin', learning AI together. And so I've done this kind of stuff in tools like Flow, but you got to do it one at a time. And then it's always like, the nice thing is, is with a Roof Claw, this session, this transcript is all saved on the hard drive. Like I can come back to this project at any time and it remembers every single word that I said. It's not lost into some ChatGPT memory that I don't have access to. I, I can come back to this transcript at any time, and I can say, hey, I want to kick— I want to go to bed, I want to kick off this project. And so, um, yeah, this is, uh, the cool thing about a roof cloud. Mexid Gaming, how often are you doing lives, Adam? Are they on a regular schedule? These are awesome. Mexid I'm glad you think this is awesome. I, I really— I've been doing them late at night because I'm convinced that I would suck on livestream. I do a lot of webinars, but those are very controlled. Um, this is more just like me kicking it real, like just trying to see it. So it's not on a schedule yet, but I was waiting for a comment like that to tell me that I'm not a bad livestreamer, and then I'll probably do more of them. All right, so let's scan through, um, So what you just said is literally why I follow your content, why I follow you, Contest. Hey man, Roach, thank you so much, man. That means a ton, um, because I want to do stuff that's of value, right? Okay, so smart call. PromptCraft is the demo here. Audience watches. Important, honest call from the front. We've only cast Marcus. Video storyboard introduces 3 different characters, so cast Denny and Carmen now. Okay, so we could parallelize that. So All right, the raw organic content you're talking about, um, is literally why I follow you over anybody else talking tech in the roof industry. That's what makes me nervous about using AI-generated marketing content. So you can use AI-generated marketing content, but like I said at the very top of the hour, like, AI is like plastic surgery— you only notice it when it's bad. So you have to invest a little bit of time in making it good. And yeah, this has been 2 hours to generate 2 pictures, right? But really what we're doing is we're doing the prep. I can cast Marcus in 100 videos from now on. Like, I can do— and again, I make videos to speak to roofers, to hopefully earn their trust, to earn their business, to help them to make my business grow. You're going to do the same thing. Like, we have a customer doing, uh, some AI-generated videos more the conventional way, not through agents. We're doing it like, you know, by hand. Um, and it's about, uh, it's a roofer in Florida And it's basically like this guy that's kind of like a Johnny Knoxville Daredevil AI-generated guy. And all he's doing is trying to test his new metal roof. So he's like, it's like a scene of like a guy on a motorcycle and it's like vroom, vroom. And his sunglasses casting across here. And then it shows him suddenly smirking, like grinning. And then you hear him like, you see the tire and then he races off and he hits the jump and he lands and he bangs off his roof and falls off the other side. And then he runs around and he goes, haha, and it's like no dent. And it's like just a series of a bunch of ways he's trying to see if he can break his metal roof. So it's episodic in nature. So the character consistency is going to make people laugh, and they're going to be a little forgiving on the fact that like maybe the motorcycle, you know, goes from being a Kawasaki to a Suzuki or something, right? Um, what do you use for the— for the cloud AI to power Roof Cloth? So I have been testing, uh, over Opus 4.7 and GPT 5.5. I've tried everything. I've tried the Quen and Deepseek and Minimax. I've connected it to OpenRouter. I've tried it. Um, I will say that like ChatGPT 5.4 was terrible. It like— Opus 4.6 was like my favorite until GPT 5.5. Opus 4.7 is still pretty good, but like it's almost kind of flipped, you know, like chat— like Opus 4.6 like it would claw with its fingernails to get the job done. Like, like you could cut its leg off and it would just crawl to the finish line to get it done. Like, it would get the task done. Um, whereas like ChatGPT 5.4, it's like the wind would blow and it would be like, hey, did you do it? Oh, you're right, I owe you an apology here, Adam. I didn't do it. This is what I didn't get done. I gotta do this. I should have done that. Okay, do that. Then you're like, 2 minutes later, 10 minutes later, hey, so did you get that done? Oh, I didn't do it. You're like, what is wrong with you, right? And then it flipped. Like, Opus 4.7 has been kind of buggy, more so in Claude Code than in, uh, in the Roof Claw or Open Claw instance. Um, but 5.5 is a freaking super genius. So if I come over here to Vito, um, and I go /status on this guy, he's running— he should be running on 5.5, right? Oh no, I flipped him over to Anthropic. So if I go here, go /model and flip over to 5.5, browse providers, and then go OpenAI and go 5.5. So I've been using both Opus and like, I've tried all the other ones. I had Gamma 4 as a local model. I played around. Like, at the end of the day, I spent more time trying to brute force and violence my way through with a cheaper model that my time is worth more than the money in my bank. Like, I value my time more than I value my money. So I, uh, I burn tokens like a crack addict, okay? So I like don't— you know, I do a lot more with AI than most people, so I— but I just find that like I want to, I want to get it done right. I want an AI agent that works as hard as I do. And so yeah, so that's what I'm using. Um, hey man, appreciate you, Damien. I'm glad you're getting value from this. Um, so when they're super good spots for AI but I'm not sold on the marketing side besides planning. What do you mean not sold? So when there's super good spots for AI but I'm not sold the marketing side. But oh well, no, so you can, you can do the marketing. So we can actually post this to Facebook ads. So if I go down here, um, so one sec, let me, let me review this and I'll talk to you about the marketing side. Um, I thought you could not get 5.5 yet on OpenClare. You not— yeah, I'm not using OAuth, I'm using the API key. That OAuth thing. You get rate limited. It's too good to be true. So I just— I use the API key. I burn lots of tokens. I just make it do economically viable work. This is your, like, writer downer of the day, um, is that when it comes to AI, you have to think like, why do I use Whisperflow? Typing takes calories. Talking takes calories. It takes energy, right? And so you convert energy, food, and calories into the body. Then you burn those calories to input the output, right? So you're— you can either use your, your finger digits, which move, you know, I type at about 130 words a minute, but it's exhausting. When I'm lazy, I type it like 78 words a minute. I can talk consistently lazy, even with breaks and pauses and stopping to go, oh, you can still talk like 135, 144 words a minute. And so it's easier to put more effort in when typing because you need to give AI a goal and then context and then action and then output. And then there's guardrails and intention and harness and evaluation criteria and testing. And like, there's— but the, the basis of every good prompt is like a goal and then the context behind the goal and then the action you want to take and what the output's going to look like. And so you want to be good at prompting, so you have to give it context. So you talk to it and you give it lots of context, and then that creates an action and an output. So then you're converting your context into tokens, and the tokens converts into economically valuable work, and the economically valuable work turns into money. And then you should be making more money from your tokens than you are spending. So calories become context, becomes tokens, becomes money. That's the lesson. That's your writer downer of the day. Okay. Um, and so no local models hit the mark for you? No, not yet, but that's because I have not yet invested in a $15,000 machine. I'm really really hanging out because I can't get the model I want even on two DJI Spark, and I'm really hanging out to see what Apple comes out with. The fact that they let, uh, Tim Cook resign and who replaced him was a, was a, was a, uh, like a hardware guy. I think that the future of software is hardware because software is so easy now that the future of software is hardware. And I'm just waiting until June to see what Apple comes out with to see if I'm going to buy some kind of crazy machine, um, or if I'm going to buy, um, the crazy new M5 Max Mac Studio. All right, so what do we got? So Marcus's cat, so cut a truck reviewing the estimate. So that's the prompt 1. So you can see here, prompt, medias, opening phrase. So it's going to say we need to make sure that these prompts are good. So I want to make sure we have the better JSON for those. All right, so I like your prompts, but I think you need to create more of the JSON prompts that create more of the consistency of the items that are around, because you're using kind of a natural language description, but I think you need to add in the JSON descriptions. So what I think you do is I think you fire up a couple subagents. One, you produce the frame 2 for the Marcus storyline, and then another sub-agent generating Danny, and another sub-agent generating Carmen. So that way you can do that in parallel because Danny doesn't depend on Carmen, Carmen doesn't depend on Danny, and Marcus's second scene doesn't depend on Danny or Carmen. So you can fire up sub-agents of yourself and go produce that. Um, and so I think we're going to basically take Path A, but you're going to run those in parallel just like you're saying, um, and that way we can show everybody what parallelization looks like on the live. Oh yeah, like, um, Roach Roofing. Yeah, you— we've generated AI videos that, like that Italian roofing one that I showed you earlier, um, there are others there are others. Hold on one second here and remove this stop screen. So here's one that our video guy did for a roofing company in the Olympics. So if I go back and if I go Olympic— so this is a very good example. This actually ran on TV. Olympic— there we go. So this— it's just dirt during the Olympics. To us, it's the arena, the starting gun, the grind, the heat. We sweat the details so you don't have to. That's the feeling of victory, and there is only one team that brings home the gold. Team Ferrell. One call does it all. So that, that ad actually ran during a TV spot during the Olympics. So can it be done? Absolutely. Um, it is a, uh, It is a— like, it— now keep in mind, that person's a professional. Like, they know how to prompt. This is taking care of that expertise. Like, the guy who prompted that video, he used to be— like, he's one of the contractors that we use for generating AI videos for roofers. But these guys charge between $2,500 to $7,500 for a video. Like, a 30-second spot is between $2,500 and $5K, uh, 90-second spots between $5K and $7,500. Like, we've spent less than $50 tonight and our time, you know what I mean? But like, that guy, he'll say like, shots starting from a 35mm shot on a, say, Sony A7 with the D35 lens using a wide pan quick zoom, or a jump shot, or, um, like, like, like pan 30 degrees left. Like, he knows how to prompt it because he actually produced video in a studio for Outdoor TV. He did a fishing show and like a kind of like a home and home renovation show. And so he's actually worked in television production. So previously all of these skills were kind of locked to the people who actually knew how to use, use a software like this, right? Like you had to know how to open up Sorry about that. Yeah, we lost you, we lost you. Sorry about that. I, uh, I guess, uh, 2 hours and 11 minutes was the limit for how much I could, how much I could store. But yeah, the, the whole concept— where was I? That was kind of abrupt. Oh yeah, so like an expert knows how to An expert knows how to prompt, right? Like, they know how to do it. Now, with the Higgs field skill that I built into this thing, it is taking what you don't know about drafting a prompt. It's like, wide medium handheld documentary shot at pre-dawn at 5:50 AM, real tilting, aspect 16:9. Like, it's creating the prompts, and I think that it could be done even better because it was just like, I think it could be done even better, right? And so what you saw me do was tell like, hey, I want you to actually add in the JSON object as before because I want you to get really good consistency. So this is just me training it to, to know what that a proper video producer would do. And so that's just, that's just the nature of developing the skill. So in the hands of somebody who knows how to use this tool, yes, they will be faster. But you can actually, you can actually implement this. You can actually do this. I've just been yapping the whole time, figuring it out, and then I build it into a skill, and then it becomes something I can use on a repeatable basis. And I'm saving you guys a lot of this work by saying, one, give you a platform from which to use it, and then two, give you a skill to make it easy to actually get it into your hands so that you're not stuck. But the people who actually do this kind of stuff for a living, yeah, they're gonna be faster, right? Because like I said, you don't have to be a good leader, you have to be a good trainer. And so I'm trying to train it so that way you guys— like, this is what they call Cinema Studio, right? Which is where you can start to combine more pieces into this thing and then start to like actually— like, you need to know how to use this tool because it's essentially— AI is now the camera, right? AI is just now the camera. So this is our Danny, right? It's creating Danny now. And then it's probably going to create Carmen. If I go into our assets, it's probably— I don't know if it's generated her yet. So we started out with Marcus, right? And now it's building me out Danny, right? And so we're— and then eventually we're gonna get Carmen, which is the girl, right? Because they gotta have a female roofer, because I know a few of them. And so that's what you're— that's what you guys are seeing. You guys are seeing us build this out. And so what is happening now in the background is 3 Claude instances. So my agent has spawned 3 sub-agents of itself, essentially giving— it basically clones itself, and then it takes out a Post-it note, right? So it takes out a Post-it note and it says, go generate Carmen. It hands its clone this Post-it note and it says, okay, now go generate the Carmen. And then it clones itself again, and it gets this Post-it note. It says, go create Danny, and it hands it the Post-it note. And then it creates another clone of itself and says, go create Marcus scene 3 or scene 2. And so it's essentially starting to draft, it's starting to work in parallel. So everything we did tonight was to build the skill so that eventually we could parallelize or create sub-agents so that it would go and then do that work for us on our behalf, so that we have the ability to work at 3 times the speed. Like, we are now, as of this moment, working at 3 times the speed as what we were working at before, because we built the underlying skill and then we made it so it's repeatable, so they can spawn these sub-agents or clones of itself to go give it a Post-it note to do a task, come back and be like, hey, clone number 1, I finished my job, here you go, and then poof, it dies. A sub-agent dies and goes away. So hope that's— I hope that kind of helps you see what we're working on here. So if I, if I stop my screen, you'll see that here. And I'm also burning tokens just on the concept of like also making it be like our co-producer of what we're up to here. Um, but it says 3 Claudius is working. This is what Lesson's compound looked like. And we did the Marcus the hard way, figured out the model selection, the JSON gotchas, the 6 anchor curation rule, the wardrobe locks. We wrote it down. Now 3 sub-agents inherit all of it for free. When this finishes, we'll have 3 cast characters, all consistent, all trained, ready for SoulID, and you'll be able to drop any of them into the scene, yielding to receive completion events as they arrive. So basically, these are the 3 sub-agents that are now green and running. So yeah, I'm— I want to show one last thing. So I want to show like how we can post this to an ad. So I want to say, hey, since we've been live now for 2 hours and 16 minutes, including Chrome crashing because we've probably been casting for too long, um, I want you to show us generating a Facebook ad. So you can use just an image from what we've generated today, maybe that high drone shot. And I want you to write a Facebook ad that says, um, you can generate a Facebook ad that says, if you want a— like, you'll, you'll have to improve the copy here, but essentially, we built a video skill that you can connect to a Roof Claw that you can use to generate, um, videos with an agent and have it working at 3 times the speed. The full skill is live at link and give the link, can be the link to the GitHub. And it's free and it's open source and anybody can use it. But I want you to kind of create that Facebook ad and I want you to create like a carousel ad maybe of the video of some of the shots that we did today so that people could kind of see the evidence of our work. I don't know if we're actually gonna run this as an ad. I just want you to get it drafted into the Facebook ads, area so that way we can demonstrate to people what this would look like. Because when we're done the video, we would do the same thing, but I just don't think I can keep everybody around for another 3 hours while we keep working on this. So now it is switching over from its Higgs field skill, and this is kind of messy because I wouldn't normally do this in the same session that I'm doing the videos. Like, I would start— I would generate the video start a new session and then go into the Facebook ad session. But I want you guys to see kind of the full cycle of— we have the skill created, we have it all existing, we went through the storyboard process, we went through the scene creation process, we went through the character creation process, we worked through the bugs, we created a repeatable skill, then we deployed that skill and got 3 sub-agents working in, um, in, in like coordination with each other, and then we would, we would con— we would continue going on the video creation, finish that, and then we would go and upload this to a Facebook ad. So, all right, so what do we got here? So draft single image ad, we used her. Okay, so it's creating— oh, but yeah, basically it's just pulling up what our ad's gonna look like. So it's basically saying, okay, draft 1 single image ad, image the V1 opening drone shot, what We— what used to take a creative agency 6 weeks now takes an AI agent 6 hours. We built an open-source video creation skill for Roof Claw. Story, storyboard, character casting, scene generation, final cut— same discipline pros use. Tonight we live built a Super Bowl quality ad brand film in 2 hours. It's not quite exactly accurate. Free open-source built for roofers. Well, you know, ask for a refund. Um, draft 2 carousel, same copy, 5 cards, each a frame from tonight. Okay, each card links to the same GitHub URL. Copy critique, my honest read. 6 weeks, 6 hours is a punch. Built for roofers. Yeah, free and open source. One might want tightening. Same discipline, close rules. Big, could replace with something more concrete like model selection. Revised version 2 if you want it. I'm with you, go with the revised version 2 in the carousel to show off the work we did. Yep, proceed. See if we got a Carmen yet. I don't think we did. Oh, we did. Okay, so there we go. So we got her in the same, same mood. So if I'm going to go here while he's doing that, we're going to go stop share. I'm going to share our screen over here, come over to Higgsfield. So you see here, so we've got these 3 agents working parallel. So here's all the different shots of Danny, right? So Danny looks a little different. So we're going to make this a little bit bigger. Whoop, right? So you can see Danny is clearly very different. And now we're getting Carmen. So here's Danny in about a bunch of different situations. Danny's eating a burger. Now we got something wrong here with the ladder, but that's okay, right? We just don't upload this to Soul 2.0. The whole idea is that we have to get something that's, you know, 90% close, because then when we go to SoulID Character, what it does is it takes all the right stuff and puts it here and makes it so that we actually generate, generate some derivative work from all of the images that we create. Remember, this time when we go to create a character, we might actually not stop at 20 photos. We might go to the full 83. Like, I might say, hey, I want to go to the full 83 for Danny and Carmen to do even better, right? All right, so back to the Facebook ad. Still waiting on your call. Yes, you can proceed with the Facebook ad and you can use the revised version too, as well as the carousel ad. I agree with your decisions. I just want to show it. Go ahead. All right, boys, this has been one heck of a night. I hope you guys got a lot of value out of this. I'm definitely going to keep working on this. It's kind of like my weekend project is seeing if I can make this video. And so it's going to take some work, right? But again, like, I've wanted to make this video since 2023. If you guys remember at the beginning, I showed you when I originally generated that That first time when I watched that video by Infusionsoft, I said, I need to make a video like this for roofers someday. And I worked on that video. You can see I've touched that script a couple times. There was a time when I almost hired people, and the people that I almost hired to make this video who are going to use real actors, they wanted to charge me $50,000. That video that SendBar Roofing did cost them $50,000. The AI-generated videos I showed you today cost between $2,500 and $7,500. And now we've spent Less than $100. Less than $100. This is the million-dollar skill. It's because then you can actually— oh no, my token expired. Oh man. Okay, so I gotta refresh the token. One second. So I created, I created a demonstration token because I thought maybe I'm gonna end up showing it on my screen. So I'm gonna— I have to go back and, uh, and generate a new access token. So generate access token, continue, then go to my permissions. Copy. So I'm just doing this on the other screen because I don't want to show you guys obviously the permission token. pw push here, paste, push it. So I will show this. So when I use this password pusher, I pasted the token, I clicked a button that said push. It's just a text box, button says push, and then I get this link. And then I can take this link and I can, I can send it over to, um, I can basically— yeah, sorry, I can— I, I went to Password Pusher, I copy the link, I paste it into a box, I click share or push it, it gives me a link. And so then I can share that link. So like, this is what it would look like. I come in here, I paste something, I click push it, and then that link I can send over my agent and I can say, here you go, this should work, it's a new generated token for Meta ads. I actually typed like a chump. So if I go here, share screen, come back over here. Man, I wish there was a way for me to quick share or quick, quick flip. Yeah, I know, we were so close, Damien. They were so close. Okay, so what it did here, so yes, basically said, okay, generate a new token. So I generate the new token. And then it went there. Script is here. Okay, add all, pause, 5-day cap, won't spend even if accidentally activated. Okay, so now I give it a second. Okay, so pulling the token. Token is alive. Account confirmed. Adam Gardner Sand, US dollar. Status 1, active. Firing the carousel script now. So it's now generating our ads. So then we should go to get into our, our Ads Manager. Ads manager. We're going to go get into our Facebook so that way you guys can see it once it's created. So what you guys can see right now, so if I go remove, I'll split you guys over here. So this is my personal Facebook, so I don't run a lot of ads in here. Right? I ran just one in here for a voice agent, right, which wasn't, wasn't a bad deal, right? We got 600— don't love that 666, but you got that many link clicks. All right, so, so done. Real Meta ad created in your account. Now, did it, did it post to Adam Sand or did it post to campaign? Oh, there we go. See, there we go. Hahaha. So we got our first ad. So here we go. Our ad is now created. We can run this ad and we can— it's got two different creatives. It's got the single image ad, it's got the carousel ad. And so we could run this as an ad today and actually use it, and then it could monitor these ads for us. Right? Like, it can monitor the results and then change it. Be like, oh, um, I guess last lesson of the day, okay? Last lesson of the day is what you guys need to know. Like, where does this get absolutely insane? Okay, so Andrew Karpathy, AutoResearch. So this GitHub repo is called AutoResearch, okay? And what AutoResearch is— now keep in mind, it's got 78,000 stars. 78,000 likes, essentially. Okay, it was created a month— like, what, what was it great, like a couple months ago, right? Let's go pull requests. Like, 150 times it's been pull requested, 150 times it's been done, or 150 open, 211 closed. This here is a looping tool. And so I'll just come back to you guys here, but Um, what this is, is this tool will actually— it's basically 3 steps. One, it has access to read the results, to measure a KPI, click-through rate, conversion rate, cost per lead, some kind of con— some kind of metric. Ideally, the, the more frequent that metric can change, the better. Like click-through rates at the top, and the conversion rate of your landing page is the bottom, and then your cost per lead is the very bottom after that. So it can measure a KPI. That's piece 1. Piece 2, it can create a hypothesis. Like, like it could hypothesize from those results a different way, something to do better, right? 3, it has the ability to, to test its hypothesis. So think about what this has. This has the ability to manage your AI, your ads. It can read the KPI, click-through rate, cost per lead, cost per thousand impressions. It can look at a metric. Two, yes, we can see it already created a version 2 of copy right there within one chat prompt. It can create a new version of copy. It can say, you know what, I think this would work better. It has a hypothesis. And then three, it can then go and create that. It can test it. So then it basically says, ah, the ad's running, we're paying $50 a lead. Why? Because we're paying $4 a click. Why? Because we're paying $15 per 1,000 impressions. We're getting, we're getting a click-through rate of 3%. I think I could get the click-through rate to 3.5% if I adjusted the copy in this way. Okay, well, I'm going to run that ad, compare them, and then pick a winner. And then say, okay, I was right. This one was a winner. Kill the old one, run the new one. Hypothesize new one, run that. You can do 30 tests a week instead of 30 tests a year. That's where the real money happens in all of this. And so then you all of a sudden, you just get this little machine running. And this in the hands of somebody who learns this skill can be incredibly dangerous, incredibly dangerous. So this is what I want you guys to pick up is like that it's not about, you know, making whether this video— again, AI is like plastic surgery, you only notice it when it's done bad. But it was able— we're able to use it to generate a storyboard, use that storyboard to generate a set of visuals and an emotional heartbeat of a video, then use it to craft the scenes, then use it to cast the characters, then use it to create the soul, then use it to create the scenes, then use it to animate the scenes, then use it to stitch the video together, then use it to put the video on a Facebook ad, then use it to run the Facebook ad, use it to measure the Facebook ad, use it to hypothesize an improvement to the Facebook look at, and then loop over and over and over again until it gets better, and keep pumping more into the pipeline. This is the million-dollar skill. And so the last thing we're going to do today before I kill the live is we're going to update the skill. Okay, so we've probably had some more lessons learned, we've probably had some more ideas generated. I think that it'd be really important to just do one last pull request before we kill the live. Everybody here seems to have enjoyed the content tonight, so good job. I just want you to do one last pull request and then we are going to kill the live. If you have anything to say to our guests, definitely post that up as an easily identifiable blurb inside of your response and then create the PR. I'll merge it and then that way anybody can now download this with the Roof Claw and we'll keep updating it through the weekend as we finish generating the video. All right, let's see what Claudio has to say to y'all. If you guys have enjoyed this, do the thing where you like and you subscribe and you hit the little buttons and bells. Um, helps me out, makes me know that I should keep doing this stuff. Um, if you enjoyed it, I'll keep doing it. I'll keep doing it until people tell me it sucks so bad that they can't stand watching and they never show up again anymore. Until then, I'll keep trying to do this because this has been fun. This has been fun for me. I really appreciate— you guys have a lot of ways to spend your time in a day. You guys chose to give me your time. I can't thank you enough for that, um, you know, it means a lot. You know, it's just you, you try and create something and people enjoy it. It just, it feels good. So for everybody watching tonight, all right, so this is from Claudio. Tonight on Facebook Live, we built a Super Bowl quality brand film pipeline in 2 hours and 30 minutes entirely on AI agents in front of you. The full skill is now open source on GitHub. Any roof claw with a marketing soul can pull these lessons and start producing tomorrow. What you saw tonight: cast a fully trained character, generate cinematic hero frames, discovered which Higgs field models to use to do which jobs, Spawn 3 Claude sub-agents in parallel to cast Marcus, Panel 2, new characters Danny and Carmen simultaneously. Drafted a real Meta Carousel ad in paused status end-to-end through the API. Wrote it all down so your Roof Claw inherits these lessons. Here's the repo, it's free. Stay tuned through the weekend, we'll keep you updated as we finish the actual video. 3 roofers, 1 ode, 1-stop source pipeline. Thanks for watching. Right on. All right, and we have, um, our new MUR, so then it's going to create me my new lessons. Thank you everybody for watching. It's been a pleasure. We'll see you guys in the next one.
