Adam Sand teaches roofing contractors how to use AI effectively by managing context and tokens. He covers system messages, a four-part prompting framework (Goal, Context, Action, Output), storing reusable skills as markdown files, and keeping AI agents running 24/7 with heartbeats and CRON jobs. He warns that current AI subsidies won't last forever.
Chapters
All right, what's going on everybody? So this is the very first, hopefully first good live stream. I tried this before. I am literally just going to go check to see if this is going live right now, uh, to see if it actually worked because it is a bit of a pain in the ass to go live. So I don't know, uh, what anybody is, uh, I don't know if anybody's here. All I care about this time around is really just trying to see if I can get this to, uh, to work and to go live. And so this might be live on YouTube. I'm gonna go check it right now, and then we're gonna get into session and context management. So the reason that I did this live this late at night was because I literally did not want anybody to show up. So if you showed up against the, uh, against the recommendation. Okay, it looks like we are live on YouTube, so this is good. And, uh, so if you showed up against my recommendation, that is on you. I fully, I fully expect this to suck. And so if you are here, go ahead and say hello, put something in the chat, let me know that you're around. But I do not promise anything other than that I will do my very best to give you guys some information you will learn from. I can't guarantee I'm a good live streamer. I did this at 9 o'clock, literally just to test the tech, to see if it would work, to see if it would be helpful. And so we're going to get into it. We're talking today about session and context management. And so this is all about how to use AI properly. So, uh, for those who don't know who I am, I have, I have been teaching and training AI to roofers. So this is for roofers. If you're not a roofer, get the hell out, we don't want you. Um, but this is kind of for roofers, the whole purpose of it being that I've been teaching roofers how to use AI lately, and the fact is, is that all of you are burning way too many tokens and struggling to get your agents to work. And as I teach you guys how to use it, what I'm learning is you have nothing in terms of knowledge about session and context management. And so this is going to be a video that's going to teach you how to actually make your AI remember what you told it to do, what you want to do, stop hallucinating, and actually get something that is repeatable without breaking, right? Without going back and having to reteach it all the time. And so before we get into it, um, I want to let you know that it's— that what's actually happening inside your agents, right, is it's, it's not magic, right? You have to understand the plumbing behind it and understand that every decision that we make with AI is all built on do we know how to properly prompt. And if you've heard this kind of stuff before, I guarantee you you're going to learn something new because I've also taught AI to people who understand a fair bit about it and they said they learned something new, and so you might as well. I'm gonna do my best to keep up to the comments, so let's get into it. So I want you to understand how a system message works. And now this is gonna be mostly designed around things like OpenClaw or Claude Code agents, stuff like that. And when you have like an OpenClaw or a RoofClaw, if you've been buying one of our RoofClaws, shout out to all my people who bought that. I appreciate the support. It's been fun teaching you, and I hope this helps refresh you or give you something new. Because a lot of people that I took this through, I drew all of this out on a whiteboard when I was teaching them, and the one thing that everybody said was, could you send me that drawing afterwards? And so this is my way of trying to kind of fix that error and give you guys this training. And so before you— let's say you start up with Claude, you're gonna have a claude.md file. If you start up with your Roof Claw or Open Claw or Hermes Agent or any of the other agent harnesses that are out there, you're going to have some kind of configuration file. So some ones that are in a Roof Clause, the soul.md— this is a block of text that tells it who your agent is, gives it a purpose, gives it a sense of who it is, why it exists, what its rules are, right? And then tools is like, here's all the tools you have to have access to. Because the thing you need to know about AI is that when it wakes up, it has a 200 IQ. But it wakes up with amnesia, okay? And so if you think of like the movie 50 First Dates where he makes her read that journal every day, he makes her read that journal so that she remembers who she is. And that's a lot like what your configuration files mean in your agent. And so that's why we have the tools file and the identity file is like its name. So I have, Well, I have like 4 of these agents running all the time. So I have Claudio, I have Tony the Wrench Clausini, I got Sal, and I got Vito, Vito Claglione, right? The Claude father. He's the guy that builds all the roof claws with me for our customers. And so when you start a new session in AI, this is it's reading its diary. And this could be 1,000 tokens or it could be 60,000 tokens depending on what you put in there. Now, who here has an opinion about how many tokens you're allowed, like should have in a configuration file? Some people are going to say like 2,000, 3,000, 5,000, no more than 15,000. Me, I believe you should have a fair amount in there and I'll tell you why. Most of the context limits are coming up to a million. Everybody like Gemini has a million, OpenAI has a million, Opus 4.5, 4.6, and 4.7 have a million. A million token context is becoming the norm, which means if you had 60,000-page diary or 60,000-token diary, that's like 6% of your total context window. You want to have as much in there and no more than how much you need. When we're looking at our system message, this is where you start. This is just where you start, but then you send a message. You say something. To your AI. It could be 1,000 tokens, or it could be 100 tokens, it could be 7 tokens, it could be— hey, what you need to know about tokens is that the technical answer is about 3/4 of a word. For all intents and purposes, let's call a token a word. A token's a word, okay? It's not technically 100% accurate, but we're just going to roll with it. And so you say something to your AI, and now every conversation starts with that. but then it has to think about what it's gonna say back. And so it's gonna say, it's gonna reason, it's gonna say, okay, what do they say? And it's gonna check that against the system message and it's gonna figure out what tools it has access to and then it's gonna decide what to do and that all costs tokens. But then there's a response. So you might say, hey, how's the day showing? Like, how's the day starting in our CRM? And it's gonna look— it's gonna reason and think and work through your CRM, and then it's gonna respond and say, we got 5 new contacts, we got 2 deals that Mark Close lost, we got 1 new deal that's created, and 1 deal that went to contract sent. So we know how much our roofing pipeline is generated today. That's its response, and that also burns tokens. And if this feels simple to you, trust me, you're about to learn something. And so every time you go through this loop right, it's not just one message 'cause you're gonna say something again and it's gonna think and reason, but all of this is in the session and you might have token caching, but it's going to package this up and repeat and package this up and repeat. And so you're gonna have more and more and more context build up and you might build, you know, using up say a 200,000 token window. Well, now you've burned up 200,000 tokens and then you keep going and it burns 78,000 tokens and then you keep going and you're at 170,000 tokens and you're nearing the limit. And this is the thing that not a lot of people know. They might know everything I just said, they might understand this, but what they don't understand is that AI generates something, a behavior that we started to call context anxiety, all right? And so it's not broken, it's overwhelmed. And you might notice that when you're talking to AI, it starts to kind of respond a little shorter. It starts trying to like get you to get to the point. You might be trying to get me to get to the point. Well, that's what that context anxiety is. It's like saying, hey, I've been, I've been listening to you yabber on for a while and I've been giving you responses. Do you have a destination for us? Because you are reaching, you're reaching towards that context limit, and that's when that the model has what we call context anxiety. And so this is where you have to say, all right, do— am I willing to start having hallucinations? Because it's going to start forgetting what's in the middle. So if you have a favorite movie, put your favorite movie in the chat because it's always more— this is the great thing about when I do these one-on-one trainings with roofers is that they always have a favorite movie and I usually know it. And so I can relate this back to you. So if you don't, if you have a favorite movie, throw it in the chat. Let me know what your favorite movie is and I'll explain this like so it'll make sense to you. If not, I'm going to talk about the Bible. All right. So more people have read the Bible than watched any one single movie. So we're going to go with that. So in the Bible, you probably heard there's a story about Jonah and a fish and he's running away from God because he doesn't want to go to Nineveh. And if he went there, it'd be like a Jew going to the German, to the capital of Germany in the middle of the Holocaust and being like, hey, you guys got to change your ways. Probably wouldn't have worked out for him. And even if God would forgive him, he wouldn't want God to forgive him because that would be, you know, kind of sad, right? Because that was a whore. So instead he runs away and then he goes on a boat. And when he's on the boat, he's in a— he's in a boat with a bunch of people and there's a massive storm and everybody's praying to their gods. And Jonah's downstairs sleeping. He's like, how are you sleeping? And then so then they're like, pray to your God to fix this. And then they kind of figure out that it's probably him that's the cause. And he has this overwhelming guilt because he knows it's him. And so what does he do? He says that he tries to convince God to spare them, and then eventually they throw him overboard, but they don't want his blood on their souls. And so he falls into the water. A big fish eats him, probably a whale, lives inside this whale for a few days, gets spits out on a beach. Then he has kind of a revelation, gets to Nineveh. Tells them to, to, you know, be nice and that in 40 days, God's going to cleanse their city, right? So then they reform, they get better, they pray, they start acting right. And now he's pissed off about it. Like, this is roughly the story of Jonah and the fish. And you've probably heard the story as a kid, you maybe know from the Bible. Now, if this filled up your 200,000 token context window, so you were, you were going back and forth telling it the story. Right? And then it got context anxiety. If you went past the line, it would start hallucinating, it would make stuff up, it would contradict itself, it would forget early instructions. Put a 1 in the chat if you've ever had this experience before, just, just so I know that your fingers ain't broken, because I can see you there. I see the little eyeball. I know you're in the room. Put a 1 in the chat. Let me know your fingers ain't broken. This is your sign to start a new session. So in Claude Code or, or your Roof Claw, Open Claw, you can just go /new, 3 characters, clean, 4 characters and a clean slate. And this is so you can go and reset your agent and that way your conversation doesn't get what we call compacted. Because when you hit that limit, that's when you're going to start losing things. And so it's helpful to think about context as something you break into phases. And it's not just meaning that you start new sessions and hope that its memory remembers. You want to be intentional about this. And so again, if you're in the room and you're a roofer, this is— these are basics. These are basics that people don't get right. And so when you're thinking about a session, sometimes it helps to think strategically, like, what am I trying to accomplish here? And so you might break it into thirds. This is something that I tend to feel when a roofer wants to get some kind of outcome, right? Let's say creating a permanent, repeatable, skill, because that's the dream, right? If you have something you wish that AI could do for you on a repeatable basis, put it in the chat. I'd love to know more. Like, let me know, what do you guys wish that AI could do repeatedly and on, like, on a kind of, like, a routine basis, let's say? Well, first you're going to plan. You're going to decide what to build. So you're going to scope the work, and you have to get an agreement on the approach. So you, you're gonna— I'm gonna give you some prompting engineering skills here in a second. But you're going to decide what to build and you're going to scope the work and you're going to agree upon your approach. That's your way of trying to say, this is what I want to do. So you might use something like goal, count, context, action, output. And then you're going to see you're going to build the thing, right? So first you decide what you build, end your session, then you're going to implement, you're going to build the thing, and then you're going to end your session, start a new one. And then you're going to do and test the thing, you run it, break it, fix it, exercise it, catch your edge cases. So if you want to start any session, a good framework— and this is just the basis of prompt engineering, okay? There's like, you, you either live at two extremes. I find you got people who are like, all right, I'm going to get my— I, I want to go download 1,000 prompts. I want to go get the Prompt Generator Pro and I'm going to download 1,000 prompts. And you're never going to remember more than two of them, or you have no idea. And so I just want to give you where you should start, which is that you should start at goal, context, action, output. So this is what I want to do. This is how I want you to do it. What, why it's important. Action output. Action. This is what I want, how, like what I want you to actually, the action I want you to take. And this is the desired output that I'm looking for. The best way I've found to explain this, one that every roofer understands, and I'm sorry for the ladies in the room, you want to talk to AI the way you wish your wife would talk to you. So what do I mean by that? Well, if you have a wife, you will know exactly what I'm talking about in a moment. So your wife comes in and she goes, she goes, Adam, Mary and Sue, they're coming to town, we're going to the wine tasting event at Mission Hill. Mary's allergic to gluten. It's so frustrating because she can never eat anything. And like, she's always worried about touching gluten. And then Sue, she's a vegan, so she doesn't eat meat. And Mission Hill has like the greatest steak ever. But she— and she doesn't, you know, she doesn't eat it, but she gets all judgy when I eat it. So I don't know, I'm super frustrated about that. But they're coming in at 6 o'clock tomorrow and I got to get some groceries because, you know, we're going to eat at the house and then I got to get some drinks because we're going to pre-drink before we go out the next night. So going out to the club, my wife doesn't really go to the club, but, you know, that could be a story. And she's— but I got this dress and it's at the dry cleaner and I need you to pick it up. And I need it before 6 o'clock tomorrow because they're flying in from the airport and I got to go pick them up at the airport. So can you get it here by 6 o'clock tomorrow? And you're like, honey, could you have just led with, I need you to go pick up the dry cleaning? Now, why does our wife do that? Well, and you know, if I got some young AI guys in the room that don't have wives, let me teach you a little something about marriage. You probably don't do enough stuff for your wife. She probably does a lot of stuff, more stuff for you than you do for her. And that's just the facts. Ladies, if you're in the comments, definitely let me me know if you agree or not. And so what she is doing is she's telling you all the reasons why she's overstimulated, overloaded, has too much stuff to do. I see all the men dropping from the chat. Sorry, boys, but this is reality. And so all the— all those things that she's telling you is because you're like, I, I'm so overwhelmed, I got so much stuff going on, I got these things to do and I got those things to do and I got all this stuff. And I know you're gonna say you're busy because you probably got something you want to do, and I just need you to pick up the dry cleaning tomorrow at 6. Can you do that? That's how your wife will talk to you. You need to talk to AI the way you wish she talked to you. That, my friends, is goal, context, action, output. Because here's the thing that you don't know about AI that you need to: AI is a 200 IQ and wakes up with amnesia, but it can read its Bible or read its, uh, journal. It can read that at 150 words a second. Like, when you're looking at AI models, it's like the latency or the speed. You're looking at the milliseconds it can respond in, and you're looking at the rate, how many tokens per second it can deliver. And we'll remember token is a word, so it can deliver tokens at like 150 words per second. So the thing that happens in our brains is that we think that because AI absorbed that text so fast and responded with an entire blog post in seconds, we think it like somehow took a snapshot. And because it reads so fast, it's like we almost think that it read every word And then we reasoned and responded, which in a way is true. But you got to remember that it can read 150 words a second, but it reads them one word at a time. And so when you give your agent your Roof Claw, your Open Claw, your Hermes agent, your Claude Code agent, when you give it that, that long-winded sentence, like, like how that wife description, it's going to sit there and go, oh, she wants me to— she's got her friends Mary Sue coming in. So what do I got to do? Do I got to like have a shower? And then she says, oh, you know, uh, Mary's a vegan. And you're like, okay, so I guess we're not having that dinner tomorrow. And then you're like, and Susan, Sue's allergic to gluten. You're like, okay, so do you mean to like clean the house of any flour that's lying around? And then she says, and we're going to Mission Hill Winery. Like, so what do you need me to like clean the car, right? Put gas in it? Like what? Like, and then she goes, and, and we're going out and I need to get some groceries. You're like, So do you want a list? Do you want me to help you with a list? And then she's like, and, and then they're coming in at tomorrow, 6 o'clock. You're like, do you need an Uber to the airport? And so can you pick up my dry cleaning because I need to hear at 6? You're like, oh, so that's what you wanted. Well, meanwhile, AI has been burning tokens trying to figure out what it is that you want to do. And then you don't get the desired output. And that's why you want to go with the process of goal. I need the dry cleaning. Tomorrow at 6 o'clock. Here's all the reasons I'm overwhelmed. And I got all these things going on. And this is all stuff. And this is why it's got to be there at 6 o'clock. So I need you to drive to Sunrise Dry Cleaners and have the, my dress in my closet at 6. That is proper prompting. So you got to talk to AI the way you wish your wife talked to you. And so you will break this into sessions. So it might be okay, your goal is to decide on an approach of how to implement a skill. These are my business rules and constraints. Then you can discuss about what the plan is and the output. You say, I want you to create some kind of plan file, a plan markdown file, and then you get— you start a new session knowing that you've committed this to memory through the plan file that you created. Then you're going to say, okay, now I want to implement the plan. So I said, hey, we built something in another session called plan.md. And I'm hoping that you can go build that for me. Here's the context, plan.md. What I want you is write the code and no redebating, and then the output is something, and then you start a new session. Then you go into the doing and the testing. So you're saying, hey, I wanna prove that the thing we built works. The context is the build.md or the Python script or the code or the skill or whatever it is. Run it, fix what breaks, result, self-annul, fix it, make it permanent. That is your goal. So you're trying to get yourself some kind of handoff file. So at the end of session 1, you end up with some kind of plan that has the decisions, the constraints, the next steps. You get it to write something that it can work with, that understands your constraints, your intention, your, your guardrails, your evaluation criteria, your testing that you're going to do. So then you can then move to your next session and then you're going to say, hey, I want to get this turned into some kind of skill. All right, this is the thing you get a lot of you guys miss is because you don't really realize what's happening in the background because it's so fast. And frankly, you're so amazed by it that you don't see what's actually happening when you ask AI to do something. So this isn't one of your writer downers of the day, is that every time you ask from scratch, it reinvents the wheel. So if you say, hey, I want you to add John Smith to HubSpot, you go to your, go to your agent, you screenshot the message on Facebook from the lead, from Thumbtack or Porch or you have some Angie's List lead, or you, you got a business card at a, at some kind of event where you're talking to other property managers. You scan the business card, you say, hey, add this to HubSpot. It's going to go there and it's going to say, hey, I'm going to write this 1,000 lines of Python, I'm going to basically import this request, I'm going to write all this stuff out. It's going to create that code. And then on Friday you do the same thing, I want you to add Jane Doe to HubSpot. It's going to write the same 1,000 lines of code again. And this is why you want to implement skills, because skills are a combination of SOPs and directives and actions and executions built into a repeatably usable file. So you use your 3 sessions to define a skill. And so you have the add a contact skill, which knows, hey, I'm going to need these inputs. I'm going to need a name, a phone number, an email. I'm going to call the HubSpot API. It's going to return a contact ID. So then now I know who that contact is. You've tested this, you version, it's yours. This would be a very simple skill in your Roof Claw. It would build that into this skill file, a Markdown file. Now, if you don't know what a Markdown file, put a 1 in the chat. I'll explain it. If you— if the quick, the quick and dirty on it is that it is a, um, it is a 25 or 35-year-old alternative to Microsoft Word. When you see all that gobbledygook that, that AI sometimes spits out, that, that Markdown, that thing where you put a hashtag and you get a heading and you put a star on either side and you you get a bold or underline, you get italic. Like, that was because 35 years ago, 35 years ago, there was no word processor on your screen. You just had the terminal, you had that ugly screen. And so when you put a hashtag before your sentence, that was like your heading. And then the conversion happened at the printer. The printer had software that converted it, right? And so now, yes, absolutely, you're totally right. It looks like I got Chad in the chat. You're absolutely— skills are like a whole new universe for yourself because then you have repeatable actions. It's like what you used to have to pay somebody a couple of grand for, like me, to build you something inside of like Make.com or something, and that would be your repeater, right? So you could have this thing do it all the time. But the nice thing is with this, you can kind of put relative garbage in and get a similar output out. It's not, it's not so deterministic. Because you can do a business card, add this contact to HubSpot, a screenshot of something that has the information, add this contact to HubSpot, or whatever CRM— JobNimbus, AccuLynx, RoofLink, uh, Proline, whatever your CRM of choice is in the roofing industry— you push it in that way. And so the funny thing is, is that so this 30-year-old alternative to Microsoft Word has somehow come back and become the de facto language of AI. And that's just one of the funny things about AI is that this is the case. And so you repeat this skill and then in your Roof Claw, or you can build this tool.md file into your environment inside of a folder, there's lots of different ways. But then you would say in my tools.md file, I want you to reference that you have this skill. So then when you go way back to the very beginning of our conversation where I said that before you open up your session, you're gonna have a kind of system message, your claw.md or your SOUL identity tools, users, all like this. The Roof Claw has like 6 configuration files. Your Claude would have a Claude.md file. That's— you're going to have that system message load up. And immediately when the 41st Dates girl wakes up and reads her journal, or she remembers, oh, I have an add a new contact skill saved. So then you can use this skill and you can even tell it, hey, use the add a contact skill. Like in Claude, they have a pre-built skill called get— use the ask user question skill, right? And, um, and then that way— now I see someone chat here, apparently Claude really does well with XML since even more explicit within markdown. Yeah, like, I'm— you chat are a wizard, right? So you're gonna know like YAML and XML and all that kind of stuff. We're starting at the basics here, but I appreciate you being here and coming down and joining us in the slums of mediocrity. But I want to make sure that people get The— yes, the Skill Builder skill. Absolutely. For those who don't know, there's actually a Skill Builder skill. If you guys want to see that, I'll show that— show to you at the end. Um, I don't know, a live is like one of those things where the audience makes you meander. It's my first time doing a live stream, so I fully, uh, fully recognize that it's going to suck. And, uh, so if you're here giving me your time, I really appreciate it. Your engagement, sir, helps. Not because I want you to like and subscribe and hit the little bell— that doesn't do much for a small channel like mine— but more or less because when I know you're here and you're getting some kind of value from it, then it's motivating 'cause I'm in real life, I'm just standing here talking to my computer in a dark room at my house. And so yeah, we want to create these repeatable skills so it doesn't rewrite that 1,000 lines of code every single time, it just does it because the skill is prebuilt and it just deploys the skill and you can get incredibly complex skills. The out of contact is super simple, right? But this could be like reschedule all my jobs that are gonna be hit by rain today. That can also be a skill and it'll be a skill of skills. And so you get into the place of the skill flywheel where you start to build more and more and more skills like the rain day check, okay? Like syncing something to Proline. So that is one of your base lessons. Your base lesson, if you are a roofer and you are trying to use something like Roof Claw or some kind of Claude Code or Hermes or any of the agents that are out there, you want to use good session management, understanding that there is a token limit. A token is a word. You're going to hit a token limit of 200,000 or a million. As you reach that token limit, you're going to get context anxiety, and then you're going to get compacting where it removes the part in the middle where Jonah and the whale is suddenly, yep, there was a guy, he was running away from somebody, he ended up on a boat, fell into a whale, got spitter on, spit on the grass, told some people that the end of the world was coming, and they were nice. Like it would lose a lot of the really important details that actually give you the lesson, right? So you have to start a new session. So you want to build the plan, implement the plan, fix and test the plan, and then now you have some kind of final thing that is hopefully a repeatable skill. What's this? AI has started to democratize integrations if you use it correctly with skills to manage context. It's truly great. It's actually kind of psychotic, Chad. The crazy thing that I am, that I am, kidney punch 2025. I love the name. It's only because I recognize the profile picture, but like, yeah, kidney punch, you are right. When you are looking at AI today, it's actually kind of crazy. Like you start to wonder if we're going to still use the user interface anymore. I don't know. Like, did you try out the new Codex or the new Claude Code desktop app? Let me know in the chat. I'd like— the new Codex that came out today and the 5.5, like, it makes me wonder why you would even— like, I think we're just going to get to a point where we just work from one screen. I did a whole webinar about how I think the future of work is from one screen. Yeah, it's, it's absolutely nuts. Let me know what you think of it. So moving on, we want to get to the next value prop of what we're talking about here. And so, all right, when we're talking about The other problem with PDF is like, if you want to train a skill, you are probably going to want to give it some information, something you already have. And you probably bought one of those roofer books from Elizabeth, right? She has Business 411. She's got some really good PDFs that are like Roofing 101, right? It's a great— it would be a great source file to teach your— like with Roof Claw, we train it about roofing. But if you are starting out with Cloud Code or Hermes or building your own Open Claw or something like that, and you want to teach it about your industry, that's going to go into its soul.md file and its tools.md file. It's going to understand who it is and what it does. Well, if you want to give it some kind of training documents, very logical, you take a PDF and you upload it. Now here's the thing, right? When you're, you're AI, and in fact AI in general, okay, this is another one of your things you need to learn about AI. AI works in the world differently than we do. Everything about how we interact with this world was designed for people. Right? Like, there's a reason why train tracks were about the same width as a horse carriage, and the car is about the same width as the train tracks. Like, we tend to duplicate existing systems in the world. That's why humanoid robots are probably gonna do a lot more than, like, all these custom-built robots, is because the whole world around it, the whole physical world, is designed for a human-sized body, right? Well, the same thing exists with a computer. You have a mouse, which is an input device. You have a keyboard, which is an input device. You're sending commands to the computer, and then the screen displays data roughly within the range of what the human brain can experience and read and consume. And so with AI, this is your another writer downer for the day. You don't have to be a great leader, you have to be a good teacher. With AI, you don't have to be a good leader, you have to be a good teacher. Because AI is like the, the, the best people who are working with AI are like the computer scientists. Okay, Kidney Punch, he's gonna be great working with AI. Now, the second best people at working with AI are CEOs. And I'll tell you why. Because you can talk to it at 2 in the morning and 2 in the afternoon, you can talk to it for 2 minutes, or you can talk to it for 2 hours. And you can tell it to do it twice and it doesn't cry. Can you name a better employee? Because the CEO, they walk into a room and they go, hey, I want to work on that PDF later, that presentation I'm doing. Okay. All right. And then employees like, ah, Which PDF? And you're already gone. And then it walks into your office at 4:59 PM, says, okay, so I want to sit down and work at work on that thing. And now you sit there and talk to it for 2 hours. And it's like, I was off work a minute ago. And so you talked to it for 2 hours, right? And then you can, you can fire it back up at 2 in the morning when you get one of those wake up and discretionary thought, shower thought that pops you out of bed, and you can go work with it. And then you can tackle it again at 2 in the afternoon. And when it does it wrong, you tell it do it twice. So CEOs love AI, right? And it's because it takes all that pressure of all that learned skill and ability of being a leader, right? And it takes all of that stress and pressure away. And all they have to do is be action takers. And typically CEOs and business owners are action takers. If you started a roofing company, you're an action taker, right? You wandered off into the foray and start a business because you're a high D personality probably, and you just went forth and conquered. That leadership part is a skill. All that empathy and having patience for people and grace and teaching and training and leading and supporting, that takes energy. And so when we love how to work with AI, that's great. We love working with this thing, but we have to know how to teach it. And we got to teach it in a token-friendly way because what Kidney Punch said is totally real. It's like you can talk to it 2 in the morning until your usage runs out, until it says, you've reached your max limit, spend more money or wait until Friday. And you're just like, okay, here's my credit card because you're in the middle of something. So here's a way to not burn your tokens and end up finishing up your usage in the middle of something. So if you have a 40,000-character token PDF because it's got pictures and graphs and tables and formatting and all kinds of stuff that your AI doesn't need to read it, because the world around you in the computer, in the digital world, yeah, you need your HubSpot to show you no more than 20 contacts and no more than 5 pieces of data and then need to click page, you paginate, you go through the pages to read it. You need that. Meanwhile, AI can consume 4,000 contacts of information and doesn't need to see it in a nice clean table. Well, same thing goes with PDFs. It doesn't need all your fancy formatting. It doesn't need that. It'll read it in Markdown. And so you use a free tool, takes 30 seconds, just Google it. PDF to Markdown converter, very simple, very widely available free tool. And you convert that, you know, PDF is beautiful and great for you and your human eyes and your need for paragraphs and structure and line breaks and clear separation of content. The same reason I build these slides, I can't I gotta put it all on one slide. I gotta go through it like a picture book because that's how we think. That's how we learn, right? And you convert it to a PDF, you convert the PDF to a Markdown, and then you feed that into your AI. Because if you've ever used something like Scribe, right? Like if you want to teach your AI, I'll show you this if you want to. If you want to teach your AI, a great tool to use is Scribe, right? And now let's see if I can— let's see if I can do this properly. Oh, see, I'm already screwing it up. Whoop! I told you livestream would suck. So if I go to Scribe— that's the thing, right? We're allowed to be bad on livestreams. You don't have to be as polished as we do on a webinar, right? And so if I go to a tool called Scribe, and I'm going to present another screen. We're gonna go stop screen, and we're gonna go present a screen, and we're gonna go over here to share screen. Scribe. Bingo. All right, so if we're in Scribe now— I'm a business that trains roofers how to use software. We have 20 people here. We've got thousands of customers. We've got hundreds of CRMs that we've built. And so yeah, we've collected a lot of, a lot of Scribes over the days of like how to do something, right? And so you might have a permit process playbook, right? And this is 5 steps. So all Scribe is, is, is a process of going through and saying, okay, when you're in HubSpot, you're gonna go to your playbooks and you're gonna click here. And every time you click on the screen, it puts one of these little orange circles on here. And so this little orange circle is your way of saying, hey, this is where you click, follow the bouncing ball. And then you're going to go into your playbook and you're going to see that you're going to have your homeowner details filled in. Then you're going to go and like see if it's right. And then you're going to enter new permit details. And then you're going to say permit requested when you did the permit. That's the process. That's the SOP. All right. Now, your initial, your initial reaction, if you want to share this with your AI to say, hey, I'm in phase 1, here's my goal. I want you to do permit handling for me. Context. This is all the stuff of why permits are important, what permits are, why we do permits, why they know when they need to be done, what part, what deal stages we've done, who usually does them, all the context. And you're going to say, this is the action that I want you to be able to take, right? And you're going to send it this, and you want to say, how would you do that? Okay, this is the thing. If you don't know how you would get the AI, like how the AI would do it, you give a goal, you give it context, you tell what you want to be able to do, and you say, how would you do this? And then, I just wanna double-check the chat here. We're good on the chat? Everybody seeing this? Yes, perfect. Then you'd say, how would you do this? Well, your first inclination when you go to export would be to go to PDF, right? 'Cause that's how you sent it. Everybody's gonna send a nice pretty document. I don't encourage you to do that. I say go to Markdown, okay? 'Cause then you're gonna get this and look, you're gonna get the hashtag and the other ugly hashtag, all of this. This is just all Markdown. And it's gonna tell you, put these Amazon S3 storage links of where the photos are. You're gonna give it this. See how much less of words and stuff this is? Your PDF is gonna require way more tokens. All right? So you're gonna give it this PDF so you can tell it what you want to do. Does all that make sense to you people? Put a 1 in the chat. Claude in Chrome has a teach— Yes, Claude in Chrome has a teach function. So if you're in Claude in Chrome, you can literally teach it, how to do it. You can do a SCARB in there. The only thing that I don't love is that, or the thing that I'm cautioning people about, is that Claude and Anthropic are in a grudge match to win your business. And some companies have got their entire Claude account shut down because some employee was doing something they weren't supposed to be doing, and they lost all their stuff, all their skills, all their, um, all their, uh, live, uh, what do you call those, live artifacts. They lost all of their embedded and automated agents. They lost a lot of their stuff. And the same thing is probably going to happen to ChatGPT. And so I fully believe that Claude and OpenAI are building super apps that are going to basically eliminate all other apps. That's the app to end all apps. But with a roof claw, all your memories are stored inside the computer. This is why I'm surrounded by MacBooks now, because my— but the memory stored in the computer, skills are stored in the computer. And if GPT comes out with some crazy 5.5 that changes the world, You have all of your memories, all of your skills, all of your transcripts, all of the stuff you've built, all of your code, and then you go change model. You could change your CRM and you say, hey, everything that you do, all of your work, all of your— all of the actions that I want you to take, I need you to read these API documents and adjust the skills to have new endpoints. Bang, you're instantly working with a new thing. It's that quick. And so I I want you guys to learn something. I don't care what it is that you learn, but you have to learn something. And if you're going to train AI, you want to train it in a way that is token-friendly. All right, let's see, does this work? Bingo, we're back. Okay, great. And so if you convert the PDF to Markdown, it goes from 40,000 tokens to 1,000 tokens. If you could find a way to say on a $40,000 roof, if you could get your COGS down to $1,000, would you do it? You bet your ass you would. So you never freed raw PDFs, you always convert them first. This is one tiny little habit that's going to save you a ton of money and a ton of tokens. And there's a lot more, there's a lot more where this came from. And so I just want you to remember this because Right now, this context saving, these token optimizations, this is a $500,000 to $1 million skill. Not the roofing industry, not the roofing industry, but Best Buy. Okay, because we have a person on a team that we work with who does like stuff for Best Buy. That is a very, very highly valuable skill when you're dealing with millions of inquiries and millions of chatbot agents that are trying to get people to design their smart home or work out how they're going to set up all their, their Hue lights. Like when you have these agents that are doing it, they're doing it in a token-friendly way, can save them tons of money. Because here's the hard facts. A lot of people are going around right now saying, oh, 70% of companies say they aren't getting results from their AI implementations. Right? 30% are. That's like saying, oh, I played basketball and I didn't make it to the NBA. It's like, Yeah, but LeBron did, and he's loaded. So you gotta like think through that lens, right? Like, you don't think through the lens of some people don't do it right and don't get the desired results. Well, of course they don't. Not everybody's going to get the desired result. Not everybody's going to get that result. You have to learn. This is a valuable skill that you need to learn— context and session management, learning how to properly do things in a way that reduces tokens. Because I'll tell you something, We are coming up to the end of subsidized AI. I don't know how much you know about subsidized AI, but right now if you have a Cloud Max plan, all right, and you're paying the $200 a month or whatever it is these days— I think it's $200 a month. I have, I have to pay for a bunch of other people that have Cloud, so we have Cloud Teams. But in it— but if you have a Cloud Max plan, you're paying $200 a month, right? Well, just like every other software company that ever came out that you know about? Like when Amazon came out, they lost money for a bunch of years, and then Jeff Bezos owned the internet. He was the store of the internet. Or when Uber came out, like, they weren't making money right away, right? Do you think they're paying for billion-dollar data centers charging you $20 a month? No. Like, all of these tech companies start this way. They subsidize the business to get users. To get you hooked. What happens when all your agents and all your automations and all your actions and all your, all your live artifacts and all your managed agents and all your workflows and all your dashboards and all your stuff's in AI, right? And you've got all this stuff in there. You're not leaving. You're gonna become platform dependent. That's why I like GrooveKlaus, because it becomes like your personal hybrid employee that you can switch the models tomorrow if you want. You could even put Yeah, the first crack rock is free. That's— you got it exactly right, Kidney Punch, right? Is that— that first— the first one's free, the next one you got to pay for. So right now, when you pay for Claude Max for $200 a month, you're getting $4,000 worth of tokens. You are looking a gift horse in the mouth if you aren't using your extra usage and maxing out your tokens and burning them up right now, because you are getting subsidized AI. And we saw the first crack just a few weeks ago because what happened was Claude Code— or sorry, Claude— well, yeah, Claude and Anthropic, right, the company that runs Claude. They— how much you know what OpenClaw is? They came out with it. And if you're roofers, I'm assuming you haven't been up on the news as much as like me and the rest of the AI nerds out there. But this thing comes out called OpenClaw, rockets to like 300,000 likes on this thing called GitHub, which is just Facebook for developers. Okay, how many green boxes do you have? How cool are you, right? Well, this guy comes out with this thing called OpenClaw, and the first thing Claude does is says— or, you know, he called it ClaudeBot. He goes, ah, they're going, that's a little close to our intellectual property, mate. We think you're gonna have to, uh, change your name or we're gonna see it. So he's like, oh crap, I gotta do that. So he changed. So he has to drop the Twitter handle and comes out with this thing called MoltBot, because lobsters molt. So he comes with that, and then it's a bunch of crypto scammers picked it up and started scamming on the name. And that's why you saw a bunch of people with Mac Minis for a while riding a bunch of hype. Because they're, they're doing a pump and dump scam on a crypto. So then he changes the name to OpenClaw, right? So Claude kind of said— and what happened is this guy's building this thing and people are running their Claude Max plans on it, which means that they were getting unlimited tokens. Now what happened is kind of like a gym membership, right? So a gym membership, you go to Planet Fitness, everything about Planet Fitness designed to keep you fat and out of shape and not going to the gym, right? If you ever got fit and dropped the weight, they kick you out. Right? And if you start getting thin, here's free pizza Fridays. So the whole thing is it's designed to keep you out of the gym, to keep you from getting good, right? But they want your $15, and you got to pay 6 months worth of membership fees to cancel, and you got to do it once a year, and you gotta, you gotta show up with a urine sample to prove that you're you to get it done. Planet Fitness, please don't sue me. This is based on things I read on the internet. All these opinions are of my own, and I do not represent the facts or figures, but blah blah blah blah blah, don't sue me. But if everybody who's paying $15 a month to go to Planet Fitness showed up at suppertime, it would be overrun. You'd have a 30-minute wait just to pick up a dumbbell. So that's what happened to Claude, right? It's like everybody showed up to the gym all day, every day. And so Anthropic servers blew up. And all of a sudden, everybody was using the subsidy to the max. So they said, hey, you're not allowed to use a Cloud Max plan with this anymore. And they cut that off. And then, but in the middle of that, everybody switched over to Cloud. OpenAI was getting dunked on, it wasn't very good. GPT-5 was underwhelming. 5.1 was just a half measure. 5.3 wasn't much better. Everybody went over to Cloud and they tripled their revenue. and it was insane. And so then Claude's— or Anthropic's servers like blew up. And you look at the reliability and the status and this Claude down report, and everybody's complaining about it, and they're trying to— so they're giving you like discounted tokens if you work at night, like discounted gym membership if you only come in at night. They're doing all these things to try and respond to it. And it— and not to mention, they started then trying to like dump all the features of OpenClaw into Claude, and they've got a lot of them. And so If you have any understanding of subsidized AI right now, you can afford to be stupid because you're being subsidized. But eventually you won't be. So the time to learn is now. Because there will be people who get hooked on this, and they're going to jack the price. And then you're going to really wish you did this later on. And so Please, please, please listen to me when I say learn context and session management. So let's get into some stuff that's kind of RoofCloud specific that makes AI really cool and makes it live 24/7 and work for you. Okay, and so the one thing is called the heartbeat. All right, this is a regular check-in. An agent that wakes itself up, looks around, sees it has anything to do. It's— these are for quick conversational things. They don't need to be precise. This is for inbox monitoring, pipeline spot checks, looking for unread messages, checking if any— anything's on fire. That's what your heartbeat's for. So when we configure Roof Claw, we write your heartbeat so we understand, like, what is the heartbeat of a roofing company. We build the heartbeat in accordance with the purpose that you want for it. So now you have an AI that works 24 hours a day. It can be every 30 minutes, it can be every hour, but this is what you set up a heartbeat Now the second thing that you have in your RoofCloud, or, and you can build this into other things, you have managed agents now, you have, um, you have things called loops in Cloud. You, you're going to see this, this kind of stuff go permeate all over, but you're not going to know how it works. You have the cron trigger, that just stands for chronological. So this is a scheduled job, it runs at the exact time every time. So it's every Friday at 4, it's go into all of my stuff and get all my numbers for my traction for my Traction EOS Level 10 meeting, all my numbers for my accountability chart, go get that, put it in Traction, in the Traction spreadsheet for me. I gotta do that every Tuesday at noon because my meeting's at 1. All right, or every second Friday is payroll, we got to do this. Or every Tuesday, or every hour, I want— well, you would probably do a cron every hour, but you know, every 8 hours do this. Or at the end of the day, I want an end of day report and I want a morning debrief. Or every morning at 5, I want you to check the weather forecast, and I want you to look at all the jobs that I have and see where those jobs are in relation to the weather cloud, and then tell me what jobs are going to get hit by rain. And then I want you to use your job reschedule skill to reschedule those jobs, send them emails and text messages, letting them know I rescheduled. I have a guy who literally did this because he works out at 5 in the morning, and he hates that sometimes at 5 in the morning it completely throws off his workout because he has to do this. So now we built it a skill. So you can have all kinds of CRONs. And so you just got to know the difference between a heartbeat and a CRON. Your heartbeat is something that's flexible, and it's doing— it can do batch multiple checks together. It's kind of conversational. It's what keeps your agent working 24 hours a day. Your CRON is that you can give it anything on a 24/7 schedule that it can do, and it doesn't matter. You can do it once a year. You could— because it's just setting off the clock on the computer. And so finally If you want to be good at prompting, you're good with AI, you've got to have that proper prompting, right? The write me an email, like, you build me a million dollar business, don't make mistakes. Like, you can't do that. Garbage in, garbage out. Generic in, generic out. You want to learn the basis of prompt engineering, and this is the basis, this is the foundation. There's 1,000 prompts that you can go download right now. Just Google, you know, prompts. There are people who literally sold prompts for $99, right? And it was just a list of 1,000 prompts. But it— everything starts on the foundation of goal, context, action, output, right? And that means this is what I want, this is why I want it, this is the action I want you to take, this is what I want to look like when it's done. You can start to add things like guardrails and intention and harness or const— constraints or guardrails or evaluation criteria or testing, like the— how you would test, right? Like you can add stuff on, but goal, context, action, output is 4 pieces, the foundation. You start here, you learn it. So a real example could be like, goal: I want you to follow up with the customer who left a 2-star review. Context: His name is Dave. He's upset about the fascia board size. Contract said 1x8. He expected 1x10. Action: Write a follow-up email, acknowledge his frustration, explain the contract specification, offer a call. Output: Plain email, no fancy formatting, 150 words max. So fun fact, this is a great time to mention the Humanize skill. So I don't know how much you guys know about the Humanize skill. But I wish I could load two screens on here so I could have my desktop and I could have this. But there's something called the Humanize skill, and it's— there's some— there's a Wikipedia page called Signs of AI Writing. All right, and so Signs of AI Writing is all like, you know, it is the em dash, right? You have an em dash. Everybody knows I use the em dash a lot before all this happened, and I'm really pissed off because now it says I'm not allowed to use em dashes. Proper English, but Nonetheless, you write something with an em dash, people think it's fake. And so there's a Wikipedia page that has categorized all the signs, and it's continuously updated by nerds who actually want to do that for fun. And they update this with all the signs of AI writing. Like, for example, the word delve, right? The word delve use went up like 4,000% when ChatGPT came out. And so it has all of these signs of AI writing, and people keep updating it, right? Because some people want the credit for doing stuff like this, so they update. I don't want to do this. This sounds like watching paint dry. These are like the people who watch trains and videotape trains and go stop by some train station to see some 150-year-old train cross a train track. All right, like this is— there's different, every, you know, there's different strokes for different folks, right? But this here just has it all. Well, there's a Humanizer skill that you can get, and all it does is read this, compare the what it wrote to this, and then improves it until it gets under a score. So it gets scored— if it's 100 out of 100, it's got 100 signs of AI writing and it sucks. And then it'll rewrite what you told it to write until it gets under 100. You say, hey, don't send anything unless you use your Humanizer skill, and it comes in at under 25, right? And that's how you can use AI to write stuff that doesn't sound like humans, because there's a Humanizer skill. So little things that you can do that just save you a whole lot of trouble. So I hope that's one of them. So that's how you— goal, context, action, output. So I just want to kind of recap what you've learned today. So you have a system message that makes your agent yours. You want to work on that. You can use init when you start a Cloud Code session, makes a CloudMD. You can fill that with all these best practices. There's tons of information on the internet. You buy a roof claw, I'll build it for you. Session discipline, okay? One, one job per session, one goal. Because here's the thing, I know about you CEOs because I'm like you too. And Kidney Punch used to work for me, so he knows this too. Kidney Punch knows all too well about me. I have this little guy on my shoulder, and you have him too, okay? It's the CEO, little, little, little Jiminy Cricket. Other people have a conscience. You have a conscience. You got another guy on the other shoulder, and he runs the, you know, it'd be cool department. Because you'll be doing something, you'll be being productive, you'll be making something, you'll be on great track, and then you're gonna go— you have a little guy on your shoulder go, that would be cool. And then you go down a wacky-ass rabbit hole. It's a bunch of shit that has nothing to do with what you're working on, because you know, it would be cool. And so some people call it the shiny red ball. Alex Hormozi calls it the girl in the red dress. I call it the, you know, it would be cool department, because I got one that's a mile wide. And every time it's so tempting. It's so tempting. You're just like, you know what, I gotta work on this spreadsheet, I gotta get these numbers together. And then, you know, it'd be cool. The next thing you know, you're building an app. All right. And so whether you call it the shiny object syndrome, or the girl in the red dress, or the, you know, be cool department, you got to learn how to put that guy in a box. And here's my strategy. When I have something like that, we, hey, I got something for the, you know, be cool project. I want to do this. Can you add it to the, you know, be cool project tracker? And then anytime I want to find out, hey, man, I'm, whenever I'm like, kind of think, you know, I want to kind of work on something, you're like, hey, what was the thing on the, you know, be cool department tracker? We had something that we're talking about, about this. Oh, yeah, it was that. I want to dive into that. And now goal, context, action, output. Right, reference that file, start it. So then you get session discipline, right? One job per session, /new when needed. You can use /status or /session. Every, every, you know, platform has something different. On your Roof Claw, it'll be /status, right? That'll tell you where you're at, how much context you've burned up. And then you want to build skills that compound over time. You want skills that know how to use skills. Right? You want to have skills that are a collection of skills, right? These are your tools. They go in your tools.markdown file on your Roof Claw. And you want to have an efficient document handling. Don't dump in PDFs. Dump in markdown files. Consume 1, you know, 1/40th of the context is roughly the math. It's about 30 to 40x more tokens to use a PDF than markdown. And then you want to take advantage of proactive automation. So with your Roof Claw, those are called heartbeats and crons. And you want a prompting framework that gets results. That is what I came here to talk to you guys about today. I hope this was a good use of your time. I hope the livestream worked well. I will probably be chopping this up and turning it into YouTube Shorts. So hopefully get all the writer downers. If you guys like this, if a livestream is something you enjoyed, let me know in the chat if you would like me to do something like this again. I hope to get more engagement. And I'll keep doing this kind of stuff. It's just an experiment. That's why I did 9 o'clock and I promised you it'd suck, but at least it was on time. We were done at 10. Yeah, and if you guys have any questions, I'm happy to answer them. If you guys just want to tell me whether, uh, what you guys would like me to talk about next, or what you're hoping to learn, what plagues you, what ails you, I would love to hear about it. Metal Yellow First Red. Metal— my metal yellow first red. I don't know what the hell that means. MYFD. I have to Google that. Metal yellow first red, what's that mean? Metal yellow first red. I heard a bunch of stuff about Canadian medals. Okay, so I— the Socialist Republic of Canada is controlling my thoughts right now. Oh, it's an emoji. Oh, colon on the left and right is an emoji. Mellow Yellow, First Red. Oh, what? It's not showing me anything. It's just showing me, uh, just show me the words. But I got to look at me dab. That was so millennial of me. Well, everybody, thank you so much. What happens when I click that? Oh, cool, that's how you do that. Nice. Okay, there we go. See, I officially learned something as a live streamer. That's how I do that. That's how All the guys that I watch, they make those things pop up and they're like, yeah, kitty punch, it didn't translate. All right everyone, thank you. I appreciate you all for being here. Maybe I'll, I'll wake up to a bunch of comments somewhere else that didn't come through on the machine, but this is awesome. It's my first attempt. I hope you guys enjoyed it. We'll see you on the next one. Bye-bye.
