🚀 Ready to grow your home service business with AI and marketing that actually works? Book a free strategy call → https://www.contractormarketingpros.com/calendar

Most contractors think AI means a chatbot that writes emails. Adam Sand is here to show you what it actually looks like when AI runs your operations — and he does it live, on screen, in under an hour, for less than $50.

Adam is the CEO of Roofing Business Partner, a Diamond HubSpot Partner that has driven over $500 million in revenue for home service companies, and the creator of RoofClaw, an AI agent system that runs on physical hardware sitting on your desk. He built RoofClaw using RoofClaw — it created the website, runs the blog, manages cold outreach campaigns, confirms payments, and onboards new customers. It is not a concept. It is a working business in a box.

In this episode you will learn:

A full-time employee at a three-location roofing company was paid a six-figure salary to check the weather and reschedule jobs. RoofClaw replaced that function for $3 a day — and freed the employee to grow into a production manager role as the company opens a fourth location.

RoofClaw is not a SaaS subscription. It is open-source software running on a physical Apple silicon machine that you own. No vendor lock-in. No company owns your memories. When a better AI model releases, you swap the API key and your entire company context comes with you.

The agent learned the complete API documentation for HubSpot and Zuper, connected to both platforms, and set up two automated cron workflows in 55 minutes for under $50 in token costs — no Make.com, no Zapier, no node configuration required.

Converting a training document from PDF to Markdown before giving it to AI reduces the token cost by 97.5 percent. A 40,000-token PDF becomes a 1,000-token Markdown file with the same information. This single change can make AI financially viable for small businesses that are currently burning through budget unnecessarily.

The weather rescheduling workflow connects to the Open Weather API, converts job addresses to latitude and longitude, checks rain probability against your threshold, reschedules replacement jobs, texts homeowners and crews, and moves repair jobs to post-rain windows — automatically, every morning.

The review generation workflow reads every email from the job thread, identifies the customer's specific roofing material and any upgrades they selected, pulls a photo from CompanyCam, writes a personalized email and a shorter text version, sends both, and sets a follow-up reminder if no review appears within 48 hours.

Whisperflow voice input lets you dictate instructions to your agent at 135 words per minute with proper formatting — including corrections. Instead of typing commands, you talk to your agent the way you would talk to an employee.

Password Pusher sends API keys to your agent through expiring links that never appear in the chat transcript. If your system is ever compromised, those keys cannot be extracted from the agent's memory.

The GOAL framework for talking to AI: give the goal first, then the context, then the action required, then the expected output. Leading with context the way most people talk buries the actual request and causes the agent to branch off in the wrong direction and burn tokens.

Working with AI requires being a good trainer, not a good leader. You cannot describe how to do something the way you would explain it to a human. You have to give the agent the business primitives — the thresholds, the ranges, the tacit knowledge behind every decision.

This one is for the home service business owner who has been meaning to figure out AI for six months and wants to see exactly what it looks like in practice, not in theory.

What's up guys? This is Mauricio Cardinal. I'm with Adam Sand. Welcome to the show.

Thanks, man. Good to be here. Thanks for the for bringing me on.

Yeah, thanks for coming back. I'm excited to talk to you. you're you're the f one of the first guests that that was on the show initially. And I I thought the show went well back when I started the podcast about eight months ago. And it's a lot of interest in this topic. So the reason why I had brought you on is I want to see all the cool stuff that you're building, especially with roof claw, which is a really cool system and and kind of tool that you build for for businesses, especially in the roofing space. and I wanted to just kind of

give a an actual inside look of how it works. And Adam's gonna show you like step by step of how to build it for your business and to give the reasons why you should have something like this in your business. So Adam, I want you to take it away.

Yeah, for sure. So for anybody who's watching, we're going to talk about Roof Claw, which is essentially my custom build of an open claw and a Hermes agent, which is open source software. It's free that you put on your own machine. And then when you have this, it essentially becomes a completely vendor lock-in-less tool that allows you to integrate all of your software and build a second brain for your business that not only you can interact with, but your team can interact with.

And so if you've used something like Claude or ChatGPT, you have vendor lock-in. You have to use them. They have your memories. They decide what tools you get. They decide how you interact with those tools. With this, you can connect anything you want. And if there's a crazy new model that comes out or a crazy new software that you want to connect into it, you have complete control over that and you can switch to whichever model you want or use multiple models in the same machine. So the idea is then.

Some common use cases that we've had is someone bought one and treated it like a marketing person. And so it was generating videos with Higgsfield, running Facebook ads, you know, scanning their website using isitagentready.com to like make their website, you know, confirm to the new standards for agents. Another person who bought it, they had three production managers in three markets, and they had one guy whose only job is to check, like a fourth guy whose primary job is just waking up every morning and seeing if there's any rain in the schedule.

And then rescheduling jobs and crews and material deliveries. And if there's no rain, he doesn't work. So he gets paid a pretty good salary to work three days a week. And if they and he has nowhere to grow, but if the company grows, he can move into becoming the fourth production manager and they could offboard this clicky-clacky toggle tax job that's not fulfilling and only needs to be done maybe two days or three days a week. So those are two kind of completely different examples. But we've got about 30 of them out there in the world, and some customers are on their second. So

it's been really incre really amazing to see what people have done with this thing once we get it in their hands.

Before before we get into that, Adam, can you explain first what is Hermes Agent? I don't think people don't know. And then I don't know if you're gonna get into it now. Or and also with Higgs field, 'cause I don't I think people a lot of people don't know as well.

Yeah. Yeah.

Yeah, so a lot of people are like the world kind of tilted on its axis on around January. There's a guy named Peter Steinberger who launched a software called OpenClaw and it's free. And the technical definition that was kind of birthed out of this was an agent harness. So you heard about like AI, and then you heard about agents, and then you heard about agent harnesses. And so open claw is that it's an agent harness, and basically it means that an agent.

Was a large language model connected to tools with memory, right? While an agent harness brings purpose and brings skills and plugins and all these abilities. And so if you and it operates at the terminal level of a computer, and I'm gonna show you what the terminal is, but essentially every software in your computer is essentially a graphic user interface on top of what runs in the terminal. Like you've all heard that at the base, a computer is just little transistors going zero and one.

millions or billions of times per second. So at the terminal level, it has control over your computer. So anything a person can do at a computer, this thing can do because your mouse and your keyboard are just input devices of terminal commands. When you move your mouse six inches up, that little laser in there is just saying move the little cursor up, you know, 700 pixels. So if a person at a computer can do it, an agent harness at the terminal level of your computer

can do it, which means that essentially you can treat these things like co-pilots or hybrid workforce that you can bring into your company to do a job that a human would do. And then it's just a matter of you learning some of the skills that we're going to teach you today, which is how you can turn cot like calories, your human energy, into context for AI that can turn into meaningful economical work, economically valuable work that turns into revenue and profit. And

A lot of what I PC people doing online is like AI theater. I have this like kind of colorful term of like jerking off to set up porn. I or and it's it's colorful and I hate saying it, but it it's what a lot of people are doing right now is they're just setting things up and they want mission controls and dashboards and and they want to like replicate a 30 million or a 30 billion dollar software program. But what you're really what actually makes your rep your business get better is when you can increase capacity.

without increasing head count.

Yeah, that's very important. I mean, that's the whole point of this. A lot of people I've seen just build tools that have no business purpose at the end the day. And and the way I look at it, just to give you context, and I mean maybe you agree with this, is that what I try to do with my team is is try to take all the tasks they're doing and and use some type of automation or workflow or AI to automate it.

Right. So some things might take twenty minutes, thirty minutes before it it wouldn't take a long time. But if you start automating each piece, thirty minutes here, ten minutes there, you're saving time and you're increasing the output. That's the way I look at it.

Yeah, you're the whole idea is that the world was largely built around things that came before it. The analogy I use is train tracks were the width that they were because horse and carriage was the width width was that width. It made sense. And then when we design cars, why are they not three wide or four wide? It's because a car is roughly the same width as a train car. So everything on a computer is essentially you interacting with data that's got

Five columns and up to 50 rows. And there's some schematic of that that's human readable, human digestible. But a computer can move a lot faster and can consume a lot more data. And it doesn't need to switch between windows because this app and this app and this app all have to give you some version of the five columns and the 50 rows to say interact with a bunch of data. It can switch between those windows instantly. And so when you can look at when you want to be like an AI native company.

Then you have to go down a layer a bit deeper than we're gonna go today. But is what is the business primitives? What is the tacit knowledge that makes you make decisions? So when you look at your computer screen for a for a blink of an eye, right? You see some piece of information, some context is in your head and it directs your next decision. How do you capture that and give it to the automation, give it to the skill, give it to that?

That script so that when it when it does it, it would make the same decision you would. What I what I find is a helpful way to think about this is that if you take your kids out to learn how to drive, and this is the because AI, I guess we'll start with AI doesn't require you to be a great leader. You can talk to it for two minutes or two hours. You can talk to it at two in the morning and two in the afternoon. You can tell it to do something twice and it doesn't cry. It's every favorite, every CEO's favorite employee. But

You don't have to be a good leader, but you have to be a good trainer. So, what do I mean by that? Well, if you are taking your kid out to drive and you take him out to a parking lot and you say, Okay, we're gonna start at this end of the parking lot. I want you to accelerate to like 30 miles an hour and stop at the stop sign at the other end of the parking lot. Well, your kid's gonna first hit the gas pedal too hard, then too slow, then they're gonna get up to 30 and they're gonna hit the brake too hard and jerk the car to a stop, and then they're gonna then they're gonna not push hard enough and then slam on it later. And it's because

You can't yourself describe it. Like describe how you like you don't say, okay, I want you to apply 30% throttle for 38 or for 13 seconds, and then apply 13 pounds of brake force for six seconds, gradually increasing to 18 pounds of brake force, and then just as you get to one mile an hour, let off the brake and then apply it firmly within three feet of the stop sign. Like that's not how you teach your kid how to drive, but that's how you have to teach AI. You have to give it these little business primitives.

Because those are the little gaps where you find AI breaks. And so we're gonna give you the foundation to get started today. But when people when 70 when they say things like 70% of companies are not getting a return from their AI projects, well, that's because they don't do this. 30% of them are, and they're having record profits and record outcomes and disrupting industries.

That's like saying, I grabbed a basketball and I didn't join the NBA. It's like a basketball in the hand of LeBron James is worth more than in the hands of the average person. That doesn't make a basketball not a money making tool. It just means that some people know how to turn a basketball into money and others don't.

So, yeah, so the so the intention today is I'm gonna walk everybody through, and I would never do this on my main machine. I would always do this on a separate machine, but in order to screen share it, I'm gonna temporarily install one on my machine and I'm gonna walk you through the exact steps of connecting, but basically building a Hermes agent, which is just a essentially a cousin of OpenClaw on our machines. We put both on. And then we're gonna set up the Hermes agent.

And then we're gonna connect it to a roofing CRM and a roofing field service management tool. And we're gonna actually build what would be equivalent to like 30 to 60 make workflows, but we're gonna build it inside the agent so it can take these tasks and automate things on your behalf. But again, I won't be getting into the specific decision logic because every company makes decisions slightly differently. Does that make sense?

Hundred percent. Can't wait to see.

All right, so we are gonna jump in. Forgive me, I've never done it this way on a screen share before, so we're gonna try and make this make this good for you guys. So I'm gonna move this window down here just so I can get it out of the way. And then we'll share our screen. Marisa, you gotta make me one promise. If I accidentally show an API key or something, you gotta remove it in post, okay? I'll let you know just because I I'm gonna be using a live client's API keys today.

Yes, sir.

Okay, so you can go to this website or just Google Hermes Agent. It's and then the instructions are really quite simple. And this is where a lot of people get in trouble, is it's so easy to set one of these things up. So I'm gonna go into on my computer, I'm just gonna type terminal.

And we're gonna paste that.

Can can you explain can you explain what a terminal is? 'Cause I think most people will not don't know what the terminal is.

Yeah. So if if you're old like me, okay, like not quite a boomer, but not quite like a Gen Z, you remember when a computer looked like this. Like this is how computers looked. They they were just a screen with one line, and you could type and you could delete and you could hit enter and spacebar. There was no mouse, there was no keyboard, like like there was no click here, click there. It was all text and everything on a computer was a command. So if you wanted to

change folders like if you want to go into a folder you went cd and then space and then whatever that folder was because you were changing the directory that you were working in. So that's essentially how a computer used to run. Then we put a graphical user interface on top of it and essentially made it so that when you click a mouse button like when you click the mouse or when you open up a like when you click an app on your on your desktop

It runs a command that loads that software up. Or when you want to download and install a software, you would go and find the app and click on and click install and drag the little icon over. All of that is just terminal commands. Like if you want to install a software, you can like if you want to install let's see, Canva or Tailscale or Whisperflow or any any of the apps that you see on my computer.

In many cases, you can install it directly from a terminal. You don't have to go into the app store and download it. You can just run a terminal command. So that's essentially what we're going to do now. Is when we hit this, right? We're going to go to this URL and it's going to download this script. This script, install.sh, is going to take control of my computer and start installing the software. So when I hit enter, it's now going to go and do all of this. Now I'm going to delete this when we're done because I don't

put these on my main machine just because it's a pretty significant attack vector, right? If you don't know how to do proper cybersecurity stuff, some good, you know, starter stuff would be like getting your own private network, right? Through something like Tailscale, it's free for up to three users and 50 devices. And you can create your own private VPN. That's like the starting line. But this is all stuff that we build into it when we when we design it.

So all you saw there was it just going and doing a it was going and doing an install run. See, and funny enough, I have an open claw on my machine from another demo that I did. So I'm not gonna migrate it. It's saying I could migrate my open claw, but since that was just a demo one, I'm gonna say no and hit enter. So now there's two types of setup. We're gonna go through the quick setup. And again, I don't recommend that, you know.

Like I would say go through the full setup if you're doing this. But essentially this isn't

Real quick, Adam, can can you zoom in a little bit? Can you zoom a little bit? It's hard to see a little bit. Can you if there's a way to zoom in? There it is. Perfect.

Yeah, I can make it bigger.

All right. So what it's gonna do here is gonna say, hey, which model do you want to use? So you can connect, you know, and again, you can't use your mouse on here, it's all up and down. So you can connect open open AI or Claude or Gemini or whatever you do. I'm gonna do open router, and I don't normally use open router, but I am for this for a very specific reason because one of my other agents is connected to open router. Because anytime I build one of these, it comes with a temporary open router key.

That I can shut off just for testing purposes. So when they get their machine, it works. So I'm gonna click open router, it's gonna ask me for a key. So now I'm gonna go over here and I'm just gonna say, Hey Vito, which is I should mention, so this is our Slack account, and Vito is the open claw agent that helps me run the roofclaw business. So Roofclaw was actually completely built.

By Roofclaw. It built the website, it runs the X Pro portal, it runs the cold email campaigns, it updates the website and the blog, it confirms the payments, it helps me build the customers' open claws. It is a complete business in a box, essentially. I that was the whole thing was to make a meta example of what Roofclaw could do by saying I could build a whole business into a roofclaw. So I want to say, hey Vito.

Earlier we created a demo API key for the podcast that we're doing today. Can you recreate another open claw or API key and put a $30 limit on it and send me the link via password pusher?

So it's yep.

Got it. I just he's using just to 'cause I wanna give this practic practical stuff. This is he's using it's much faster, it's much easier to instead of typing, right? You're actually you're using what whisper flow to to speak to it, right? Can you talk about that? The the the yeah.

Yeah. Yeah. So whisp so Whisperflow is a tool that I use. So you can see I put 500,000 words into this thing. And I can type between 80 and 90 words per minute. And I can talk at like a rec like a fast pace at around 135 what words per minute. But sometimes I'm talking about complex stuff. So I'm like,

And that would normally break your like typical voice input on your phone. Whereas this, it'll actually properly format your words. And if you go, I want to do 13, not 14. If you change from 13 to 14, it'll just put 14. So I use Whisperflow, it's it's changed my life. So yeah, definitely want to use that. Okay. So the other thing that I want you guys to know about is password pusher. Okay. And so if I go to

Yeah. Yeah.

password pusher. I'm gonna open up a I'm gonna open it up right here.

I don't ever want to send API keys into the transcript or come out of the transcript with an agent or an agent harness because it exists as a memory on your computer forever. So that API key is there forever. If you were ever to get hacked, right, compromise your system. If it was just API key in the chat transcript, it would be there. But when I go like this and put API key and go push it, when I set what I do is I send.

This link to my agent and then it opens it, adds the API key where I wanted to add it, and then it never ends up in the transcript. And this link, right, will expire in three days and five views. I usually turn this down, right? And so that way, 45 minutes from now, if someone ever went into the transcript, this link would be expired. They would never be able to see the API key. Does that make sense?

Yes, sir.

Okay, so I have to take this API key. Now I'm gonna kill this API key later. So if any of anybody sees this, but see I I right now I I can just copy this to the clipboard and go take the API key and put it into my into my Hermes, right? So I'm gonna come over here, I copied it, right? And now if I paste it here, you can see that's the API key. And again,

I'm I know that's like super not cool to do on a podcast. I will either Mauricio blur it out, but I'm gonna delete it later. It's only got 30 bucks on it, it's not worth it. So I'm gonna come back here and just hit paste. Okay. So now I've connected this thing to a brain and it says which model do I want to use? The cool thing about open router is that it can kind of use any model you want, right? And so

I can use Deep Seek or OpenAI's keys or Quin 3.7, which is apparently pretty cool. I haven't tested it out yet. So I'm probably just gonna go with what's stable and I know works, and we'll go with like Opus 4.7 or Sonnet 4.6, right? should I explain what in out in cash means? Okay. So whenever you're looking at this screen, what you're seeing is how much it costs per million tokens in.

Yep.

And million tokens out, and then when it caches tokens. And so all that means is that a token is roughly a word. Okay. So you know it's technically three-quarters of a word, but for simplicity's sake, call it a word. Now, if you put in a big long thing that's like 4,000 words, well, then that's 4,000 tokens in. Then it's gonna think, right? It's gonna think about what you said. It might burn 40,000 tokens thinking about what your 4,000 tokens meant to give you an answer.

Then its answer is just a thousand tokens. So the in is what you give it, and the out is what you get back. And the cash means that it holds certain things in the cash so it doesn't have to reprocess it. So if it's holding a hundred thousand tokens and it's like short-term memory, it'll charge you 50 cents per million tokens of cash. So it's gonna it's the way that they found ways to save money. But you can see that the price is very huge, like open AI GPT 5.5 Pro.

You're only using this to like invent the next vaccine. Okay. Like you're not using this to write a blog. So and so go ahead.

Yeah, yeah. And it and what w one thing that also it's I mean, for in our industry it's it's probably not gonna it might be one day, but overall a lot of companies, especially more the bigger tech companies that have a lot of software developers, they're they're kind of burning through through the tokens and actually it might be sometimes not economical feasible, because if you have so many people running tokens actually might might get very expensive. So it does actually can be very expensive. I know people that are spending sometimes

you know, five K, six K a month just on AI tools because of the tokens they're using. So i you know, our purposes, sm small businesses, even larger businesses, I you're you're not gonna use a ton of tokens. I mean I'm I don't know, you might have a different opinion on that.

Yeah, this is a good this is a good opportunity for a lesson. Okay. So session and context management is huge. And so when we go to like teach a company how to or when we want to teach AI how to do something, it's not always going to use the screen, right? Like it can use a computer now. It can open up a browser and open up a software and click through a software, but there's faster ways for it to do it.

But the problem is that you as a person don't know how to say, like, I need you to hit the permit endpoint and reference the contact object. Like you don't speak computer, just like that 30 pounds of break force and then let off two feet. Like you don't speak the language. So when you're teaching AI, you might say, Hey, this is how we do permit processing in HubSpot. So you when you go here, you click on the permit up, you click on the permit playbook.

And then you review the homeowner details, and then you review the project details, and then you enter your permit details. Well, where does it get the permit details from? So you would maybe add this, get them from the mgov.ca website, blah, blah, blah, blah, blah. Right? You might add something here. But here's where the lesson comes in. So if you wanted to give this to an employee to say, hey, this is how you do permits, well, you might give them access to this, or you might export it to a PDF.

Right? And so you'll end up with something like this. And this is pretty explanatory, like a human being would benefit from seeing it this way. But and this is a very simple five-step process. But if you want to if you gave this PDF to an agent to AI, it would consume 40,000 tokens to process reading this PDF. Now, if you want to give it to your agent, you'd export it in Markdown. Now, no person

Unfortunately, this is the biggest I can make it go. No person, if you gave this to a new employee, no person wants to read this. Like, this is not how people think. But if you give this this code to a computer and say, hey, this is my process for doing permits, how would you do it? Well, it would consume a thousand tokens to learn this process. So when you hear about companies that are not having financially feasible use of AI,

It's because they're dumping things in P into PDFs and then shooting it off to AI and saying, learn this when you could just use a simple website like a PDF converter, PDF to Markdown, and save yourself literally 140th of the tokens, like two and a half percent the cost. So I hope that empowers some people to save some money. So

Mm-hmm. Awesome. Yes, sir.

for the sake of this, we can probably use Sonnet 4.6 or Opus 4.7. I'm gonna use Opus because we're be teaching it how to use software today. So we're gonna go like this, we're gonna keep it local. Now, what it's saying is set up messaging now. So, do we wanna set up messaging? The thing I like about these Asian harnesses is that we can connect other channels to it. And so I can talk to these things through Telegram. So we're gonna set that up now. So we're gonna say we want to use Telegram. So I hit space.

configure and it's gonna say get a telegram bot token and so we're gonna do that next. Where is my telegram? Right here. Okay. So we're gonna go here and we're gonna go at new bot.

sorry not at slash new bot all right we're gonna call it mauricio m-a-u e cio hopeful i did that right mauricio test bot okay

Okay, so this is my token. So I'm gonna copy this. And again, I'm gonna I'm gonna kill all this later. And now we're gonna go back to our our bot token and paste it in. So now I have to say, okay, allowed user IDs, and so I can leave it open. I'm gonna usually you would say what telegram user IDs are allowed to use this. So you can actually create like an allow list, and you can say, okay, I, Adam.

Have unlimited, unfettered access to my agent harness. But then this second list of people, which are my employees, they have a downgraded access. So they can't add tokens, they can't extract skills, they can't download customers, they can't do that. But they can, they, they can do the permit process or they can do the job rescheduling workflow. And they can do it from Telegram on a walk instead of doing it from a computer.

At their desk so they could wake up and still walk their dogs. I like that's a really good example of somebody that we actually that we actually did built one of these for. Is I had a guy who's a 53-year-old Marine. And if you want a dead, broken, or pregnant, call this guy. But if you want something done on a computer, like this guy just gets frustrated. Like when those three browser windows are open on the computer, he gets frustrated. And I was terrified when I sold this thing to him that he would hate it. But

Mm.

I trained him for three hours how to use it on a Saturday. And the next day he was stuck in the airport because the Illinois got a storm. And he was on the airplane for three hours and in the airport for nine hours. And he was he messaged me. He's Man, this is awesome. I can't thank you enough. And I was so scared that he was gonna hate it. So I was like, How's it going? So I gave him a call and he's like, Yeah, I you know, RoofLink doesn't have an endpoint to to to create an invoice. And so I asked him, Well, what can I do? And so it said, Well, I could create one in Google Sheets and it

Created invoices in Google Sheets and I invoiced like a hundred grand and I hate being a money collector, but it wrote the emails for me and it did everything. And tell me why I should cancel half my software and just use this thing. And I I say that like the concept of being able to communicate through Telegram and actually run your business from the airport was something that I didn't predict would be a use case that made a customer happy. But the fact is that like things that he didn't like doing were able to be done.

And he's a Marine. He wakes up every morning at five in the morning, all right, to do his person like his physical training, right? To go work out. And when he has to do that reschedule because of rain thing, he doesn't come up for air until 11 in the morning. He's like a one man show doing two, three roofs a week. So if he has to reschedule 14 jobs very manually as one guy, it takes him all morning. And then he misses his workout and he's ruined his day. Now he can just do it from Telegram.

In between sets if he wants, or just have it do it all automatically, like it the skill gets better over time. Does that make sense?

Hundred percent. So you have this case where basically it's you're you're hiring multiple employees on your behalf instead of just you can potentially just run a a one man operation and be a much bigger business. That's kind of the idea.

Yeah, is like this this is giving a birth to a one-person company that's crazy. The story that everybody talks about is this guy, Hildebrandt, Michael Hildebrandt, Matthew Hildebrandt, he built a company to 400 million revenue with one employee. The nearest competitor, because the second year he did 900 million. The nearest competitor is hims and hers, and they did billion dollars in sales with 2500 employees. He had him and his brother. So yeah, it's kind of a wacky business, but

You can't throw the baby out with the bathwater. His nearest competitor has 2,498 more employees than him. It's changing the game. So we're gonna say allowed user IDs, all right? So home channel where Hermes delivers crown job results. So all this means is like when it does something chronologically. So when it it causes you can say every Friday at six, I want you to calculate all the payroll hours.

Amazing. Yeah. Yeah.

Or every first of the month, I want you to calculate all the job costs or update all these statements or whatever, right? Or go through and count how many leads we got or go check Google Ads to see what our Google ad spend was and divide it by the number of appointments we scheduled because it can bing off of all these systems. When it delivers that result, where is it going to do? And we can set this later, but you would say probably Telegram, right? So install the gateway, you just hit yes.

Start the service now, yes. Okay, it's installed. All right. So we're going so now if I wanna do this, so I'm gonna, I'm gonna create a, I'm gonna have to close this to create a new window. This is like your instructions, but I'm gonna create a new window.

Okay, I'll make it bigger again.

Hope that the people watching this are digging this because I'm trying to go a little bit deeper than stuff you might see on YouTube. All right. So we're gonna load up Hermes. So this thing's this thing's alive now. Okay. I can leave this window closed, I can talk to it right here, but it's alive now. So if I go back to my Telegram and click on this t.me link, we're gonna see if it's alive. Okay, so here's my pairing code.

So I can just go like this, go here and paste it in here.

So it's approving my telegram. Okay, so now I can go look at this.

Hey, are you alive? Or I'm gonna say awake. It's not alive.

All right, so we'll just set home. Actually, we bother. Okay, so it's here, it's what's up. Now, this thing has no brain yet. It has no purpose. Okay, it's just an empty software. So this is where we teach it what we want it to do or who we want it to be. Now, when I build these, we we do a lot bigger, more intense build. Like we actually each of these, like there's a soul.md and a memory.md.md means markdown. So there's a soul.

Mart file and an identity file and a tools file and a workspace file and a users file and all this stuff. That's all the work that we do so that this this thing shows up ready to work, pre-trained as an employee. All you have to do is give it, it knows that it lives in a car and it drives and it hits the brakes and it turns left and turns right, but you got to tell it how fast you want it to hit the brake, how hard you want to hit the brake. Like you got to give it some of that tacit knowledge, but it's gonna understand how your CRM works. And so we're gonna walk through that now.

Yeah. And how do you build that that that that that how do you build that memory? How do you build that is specific for each company, right? Each each company is going to be different, right? So how do you build that though?

Organization.

Yeah. So we so we have a yeah. So when somebody buys the roofclaw, they get sent an intake form and it has a bunch of stuff like, you know, what's the biggest bottleneck in your business? If you could wave a magic wand and fix one thing, what would it be? You know, what would, you know, it what what's the name, right? How much pushback do you want from this thing? Do you want it to be really brief? Do you want it to be really verbose? You know, what job do you want it to do? What softwares do you have?

You know, you know, a little bit about their business, you know, residential, commercial, that kind of stuff, materials they install, you know, what's their market, their city, right? These kinds of things. They're Google My Business. Little actually, I think we just researched that on our own now. But like we get a bunch of information, then we do a one-hour intake call and we find out like what would 30 days of using this be? Like it would be a win if it did this after 30 days. Okay, what does 90 days look like? And like l talking through the deeper nuances of the problem. So

That rain day rescheduling thing, we found out that it's because they have one employee who basically can't rise up through the ranks because he's stuck doing this job and he's got to do this. And they don't want to promote above him, but they need this part of like the weather rescheduling to go away so that when they expand, they can promote him from production coordinator to production manager. But right now, this guy only works three days a week, maybe two days a week. Like we get into the nuance of the problem. Then I take that one-hour intake call and that intake form and then

I have a whole ton of business primitive knowledge on the roofing industry. We do a bunch of deep research on them, their company, their org chart, everything that they give us in the intake process. And then we use AI to construct all of these files. And then we do a bunch of cybersecurity and boring stuff that's like good best practices. But then we build that onto the machine, ship it out to them, and then we spend three hours training them.

How do you how do you connect all the different tools they have in their business?

We're gonna we're gonna do that right now. So I'm gonna say, I'm gonna go here and say, your job is to monitor and take actions inside of two softwares that I'm gonna connect you to. One of them is HubSpot, which is our roofing company's CRM. We're a commercial roofer, and you're gonna learn more about us as you go. But we use HubSpot for our main CRM and marketing tool.

As well as it's our jobs are managed as tickets. We have custom objects like building, and we have a software called Company Cam that used to we used to take pictures. And then we also use a tool called Zuper, which is our field service management tool. This is where we do our inspections, take our job photos, it's where we do our estimates, it's how we schedule our jobs, we do our work orders and our purchase orders out of it, as well as maintain our inventory in it.

And so, what I need to do is give you both of the API keys for those softwares, and then I'll provide you with the API documentation. And what I want you to do first is smoke test the API key to make sure you have access, and then I'll give you the API documentation so you can review it. And then we're gonna go through the process of building some skills together. But first, I just want to make sure you understand what we're about to do. And if you have any questions.

Use the ask user question feature to get a little more clarity.

So again, calories, energy into context, the prompt, which I tend to follow a framework of like goal, context, action, output. And what is that? What that translates into is like you want to talk to AI the way you wish your wife talked to you. So when your wife talks, and my trust me, my wife hates when I use this story, but it works. When your wife talks to you, she'll come and say,

The girls are coming in on Friday. we're going to go to the wine tasting event on Sunday. It's at the Mission Hill Winery. So they're flying in at six. Mary, she's like she she hates meat, so she's like a vegan or whatever. And then Sue, she is gluten-free. And so I gotta go and pick up some food for them and then gotta make make meals. But I need to, we gotta be ready on Sunday because the wine event is at six. But I gotta get them from the airport at six on Friday. But the event's on Sunday at six.

And I I need you to pick up the dry cleaning.

And you're like, could you have just led with the dry cleaning part? I didn't need to hear all the other stuff, right? And so when you're when you're talking to AI, what happened? What the mistake that we make with AI is that because it can read so fast, it reads at 150 words per second. It's almost like it takes a picture of our words and then instantly does what we want. Like that's how we experience it because it reads faster than our brains can comprehend. But it is still reading your instructions one word at a time.

Yeah, yeah.

And so when we talk about token use, right? Over using too many tokens and not being viable, it's because if AI was hearing that story from your wife, it would go, you're going to the wine tasting event. Do you need me to give you wine pairing ideas? it's at six o'clock on Sunday. Do you need a reminder? you had to pick up Mary from the airport. Do you need me to look up your emails from when she said she was coming in?

she you she's a vegan. Do you need recipes? and your other friend's gluten-free. Do you need cleaning instructions to make sure there's no cross-contamination in your house? you're gonna be going at six o'clock on Sunday. Do you need directions? you need to pick up the dry cleaning. what's the ticket and which dry cleaner? Like, you know what I mean? Like, if you give the goal up front, because what your wife's doing is she's saying,

All this stuff is overwhelming me, and I have all these things going on, but you're a busy business owner and you're gonna tell me you're busy. So I need to show you how overwhelmed I am so that you'll just do it when I ask. And every husband in the room knows what that experience is like. And so we just have to recognize that when we talk to AI, we want to give it a goal. We want to say, This is my goal. I need you to pick up the dry cleaning. Then you can give it all the random context about all the stuff that's going on in your life and everything like that.

Because then it'll understand, okay, there's a lot of things going on and it needs to be the goal. So the context is all the stuff. Then the action you need to take is like pick up the dry cleaning and have it in the closet by six o'clock. So that's the output. Closet by six o'clock. So goal, context, action, output. So now we got our response back. And so it said, you got it, here's my understanding of what we're about to do. I'm gonna give them two API keys, it's gonna smoke test, I'm gonna give it the API documentations, and we're gonna iterate, iteratively build out skills together.

So, what a skill is, is it's the combination of a directive and an execution. A dire think of it like a company, right? You have a director, and then you have a manager, and then you have an employee. So a directive is like the SOP, right? It's the how we do things, the why we do things, the what it takes to do those two things, right? This the the bit of the business primitives, the the ranges of things, right? And then the execution is the employee. Now

The employee is the one that actually does the thing. And in the middle, you have a manager that orchestrates accomplishing the SOP, the goals, the directives of the business. So the agent harness lives as the orchestrator, the manager in the middle. The execution is scripts. So it doesn't have to use AI to write the same thousand lines of code every single time you want to add a customer. It just creates a Python script once, and you don't have to know how to code.

It's gonna do all that for you, but it writes the Python script once. And then every time you invoke that skill by saying, Hey, I want you to add a contact, or I want you to create a new Google ad or I want you to run our month ends, our month-end report, it just has a script that it does by itself. And it does it in a course. So when you combine directives and executions, you get scripts, and that becomes a skill. Okay. And so here we go, it understands a little bit about what we're gonna do.

And I'm not giving it the company cam API key. So I already loaded up the password pusher link for this. And so I'm gonna get that key. I tested to make sure they worked before we came in. So hopefully the key hasn't expired. Okay, here is the API keys as well as the instructions. Do not echo them back in the chat, but I want you to add them to your credentials and your.env file so that they're stored. And then I want you to smoke test to make sure that.

These API keys work and that you can successfully access the softwares. The region for Zuber is also in.

I'll paste in the link. Now, I don't know if this thing has browser use activated yet. So because I didn't go through the full setup. But when you do the full setup, you can usually enable a plugin that allows it to open up a browser and view things. And so I don't know, I because I I usually go through the full setup, I skipped a little bit. we're gonna see if it figures it out. But if it doesn't, well then we'll have to do something different.

Yeah, yeah.

Can you explain what are an E and V files?

Yeah, so a.env file is basically the environment variables that contain the API keys. And so an API key is just like it sounds it's a key that lets you get access into a software. So you say, hey, I wanna do, I wanna access this part of a web page. Like so it's like, you know, aculinks.api.aculinx.com slash get slash contact. And that's like you want to get a contact, right?

Well, when it tries to access that part of the of the web page, it's gonna say, What's your password? And you're gonna say, Here's my API key. And then it goes, okay, got it. That's the API key. And so then it gets, and then it can get the contact. And so an API key is just what gives you permission to do that. Is again for all of my super technical people in the room, I know I'm oversimplifying things greatly and just have patience for me. I'm talking to roofers that I don't know, and I'm just trying to make this easy for them. I know that like

They would need to do more stuff. Okay. All my nerd friends just cut me some slack today. All right. So I don't see a building object schema list. Only build, deposit, change, order, commission, marketing, expense, warranty, worth flagging. Okay. So it might have a listing.

Okay, and it's connected to Zuper. So now what I need to do is I want this thing to learn the API document. So it's got it here. And so I'm gonna go up here and I'm gonna just type Zuper API documentation. Getting started with Zuper API. And we're gonna go to, it's got our API reference. Yes, we're in the right place.

And then we're gonna go back.

And again, I can be doing this in Telegram as well. So we're gonna do some other stuff in Telegram while we're doing this. Okay, so here is the API key, or no, sorry, here is the API documentation.

for a Zuber.

So it's gonna get working. Now keep in mind we also have Telegram. So it's gonna say, can you ask me some question? Actually, we'll get a little bit harder to ask. Can you research my company and get to know my market, my company, and my products so that so that you can get a cursory understanding of.

Who we are, what we do, who we serve, where we serve, and how we present ourselves online, our branding, our colors, comment statements. You can look us up on Google, find our Google reviews, learn about what makes our customers happy. I just want you to do some preliminary research on me as a company. And then if there's anything that conflicts or has some kind of confusing point, I want you to ask me a series of questions and just get to know me and my company.

If you notice what he's doing, yeah, he's he's basically training training it. That's what you're doing. You're training the the actual like your your new employee, you're onboarding a new employee, you're telling them your core values, your SOPs, what your role is, what your you know what what what you're responsible for, things like that. That's what you're basically doing.

Yeah. And so like this is like the drastically oversimplified version of like what we do when we go to build it. We have a series of very structured prompts and very structured expected outputs that turns it into a controlled number of characters and and so it doesn't bloat it with 60,000 tokens of useless information and stuff like that. But like this is the broad oversimplification of how we would

How we would when we go to build something for some. But if you're doing this yourself and you just want to do it, you know, like that, like this would get you by. It's just, you know, you would eventually have to like learn more and and and clean it up, right? Because it's going to bloat the memory. It's going to do all kinds of stuff that make it kind of messy, right? And so what it so now what it's doing inside of this one, and again, I could be doing this anywhere. I'm just

I have two sessions running. You can see here I've used up 58,000 tokens of my million tokens so far working on this project. There's so much that you can do, such as slash commands, right? And if you use a slash command like goal, you can set a project that'll work for days. I've had one run, well, that's not true. I've had a Hermes, or actually, I've had an open claw.

Run a project for days. It's actually collecting Facebook comments right now. Basically, it's like it's collecting comments about some stuff that I've been kind of just seeing what the public says about these things. And it's like collecting it to essentially put together some brand messaging, right? It's putting together like, how do you speak in the way that the roofers speak about this particular topic? It's collecting all the comments.

So I don't have to go and copy paste and remember, this guy commented on that thread. So I gotta copy that the the click another notification and that same that comment was in the same thread as the other one and and and then this notification got lost. It makes sure that I don't lose any of the whatever it's at now, seven hundred and seven hundred eight hundred and fifty comments. Yeah.

Yeah. I've been I've been seeing your posts on yeah, I've been seeing your posts on different groups. so it's collecting it's is it giving you like a summary of what people are saying?

Yeah, it's it's basically breaking it into like mutually exclusive, collectively exhaustive points, right? Like what what are all the like threads? You know, everybody's got their different version. Like there's like the comic book anti PE story, right? Where it's like PE bad, but it's like, why? Because it's bad, because they change quality. Okay, how? I don't know. I just don't like them. Okay, you know, like they're like there's that group. And then there's another group of people who are like, I worked at a

Company that got bought by PE, and this is the specific ch thing that they changed. So it's like there I learned both about a second category of people, but then I also collected like a specific thing that PE does wrong or that's perceived to be wrong when they buy a roofing company. And so for so it it's doing all the hard work of collecting that. And then you can see here it's like it's a cron job, right? So remember I said crons, right?

So then it goes and runs the cron. It's got private equity showdown, private equity kills culture, private equity cold calling roofers, private equity acquisition window, private equity quality drop, Xactimate AI agent, owner culture, chaos absorption. So these are the posts that I've run, and it's captured 850 comments across 79 post surfaces. It scraped 13 and found 13 new comments, and it does this every 180 minutes. So basically just and it's been running for days on this project. And that's a

Yeah. And that's that's a lot of feedback. That's eight hundred and fifty comments is is a lot of it's a lot of int I say a lot of intelligence that you gather. That you have you have a good sentiment of what's going on.

Yeah.

Exactly. Okay, so need a different approach. Let me kill the listener and clipboard seeing this complete flat index. Let me save it to disk. And is it still working? It's preparing right file. It is still working. All right, let's go back to this one.

that's Marcus. Let's go. Mauricio Testbot. Okay. So got all services. Now let me grab visual brand info, capsule goodness. Okay, so it's it's going through and it's collecting all this information about the company and it's gonna save this kind of thing. I'm gonna go allow once. It's gonna it's gonna save all of this information. So that's a this is a good time to say, like, why do I like putting this on a physical machine? Well, when you go into Claude, right? So if you go into like your

If you go into your Claude account, right? If I'll just open up a tab here and go to my clot account, and if I go into my memories, so I go like settings and I go to somewhere in here, I don't even yeah, so instructions, but then there's memories here somewhere. I don't even look at this anymore. Account, capabilities, connections, where is my memories?

It's in it's in the in the project site, organization settings.

Is it an organization? 'Cause I thought it was gonna only be my memories in there.

They don't prep

Okay, but there is.

I think I think if you go to if you go to the left hand side you go to Yeah, I think I don't know, it's it has changed I think because before you see go to go to settings and you'll see it there.

Yeah, it's like a it's usually there's like a thing in here says memory, but I know there's a place to export it.

It's it's

Yeah, you can you can see it on each chat and that that's that's where you're gonna see it too as well.

Yeah, so so so in here there's a memory and you can yeah, see you can import and export. There you go, view and manage memory. And so this is like all the memory. That's what I've got. That's my memory. When you're in when you have your own system, right, like like this, I'll give you an example of what that looks like. So this is how it thinks about memory. can you see that? There we go. So every single chat, every single conversation, right?

Is saved and then turned into a markdown file that it remembers everything we talked about, every transcript, every script it wrote, every skill, every task, every single thing we've ever done, every single session is turned into a memory and kept on the computer. So I own the memories, not Claude, and then I can switch to a new model. Like if Deep Seek comes out with an amazing model that's a dollar per million tokens in.

And three dollars per million tokens out, and it's ten one tenth the cost. I can just switch the API key and boom, my life doesn't change at all.

Awesome.

Jacob actually worked at the company and purchased it from his former boss. So this is a long-term legacy where the employee culture has passed down, maybe not through a family generation, but it's still a family that owns it, but it's passed down from an actual employee who's carrying the legacy on. Might be something that I would say back, and it's gonna add that to its memory. So going back to our little Hermes here, okay. So it has gathered.

Through this, god kill, nothing to do. Ready for Hubspot? Okay, it says ready for the HubSpot API docs. So now we're gonna go get a HubSpot API docs. HubSpot API documentation.

Now their API is massive and so we have to this is the right one, blog CMS, CRM, yeah, okay, that's everything. Where do I go here?

We're gonna give it these. Here are the Hobswat ones. Now, this is gonna get cool in a second because what we've done is like let's recap where we are. So we've given it the keys to the kingdom. All right. It's on now on in Telegram, we've told it who we are, right? And so we've kind of said, okay, that I want, like I'm gonna say, can you commit this to your memory? Instantiate the memory dot markdown file if you haven't already.

And make sure this becomes part of your soul.md so you understand who you are. You're here to help me run this business. And that's again drastically oversimplified, but run this business is a terrible prompt. Okay. Like, what do you mean by run this business? But I'm just I'm speed running a build that usually takes me days. Okay. And I'm trying to do it in an hour here. And so we've we've told it who we are as a company. It's got some preliminary understanding from

Mm-hmm.

A blank empty slate. It's a 200 IQ super genius, all right? But it wakes up with amnesia every day. As soon as this gets to 100%, I gotta start a new session. But because it'll commit it to memory, it's gonna be like that movie 50 First Dates, where it wakes up and reads its journal and it starts and it knows instantly who it is again. It fakes having a personality. And so now what it's doing is it's consuming these API documents, and as soon as this part's done.

We're gonna get it to start building some skills. So we're gonna say, hey, sometimes information exists in Zuper, sometimes it exists in HubSpot. And I wanna be able to do certain things, right? And this is where I run the risk of like I'm working in a real live HubSpot account. So we're gonna try and demonstrate some mock examples and maybe create some like test leads that I can remember to go back into this client's HubSpot account and delete in a minute.

But think about that 101,087 endpoints captured. So if you don't know what an endpoint is, okay, so if we go down here to CRM, okay, and so if we go down to objects, okay, so a contact is an object, right? Like a contact is a person. So a building is a custom object that we create. It's a separate, like now they call it a listing, right? So they call it a listing, but essentially a listing is a building, right?

this is great stuff, huh? Yeah.

But a contact is a person. And so an endpoint is each one of these. So create a contact is an endpoint. Retrieve a contact is an endpoint. Retrieve contacts, plural, is an endpoint. Merge two contacts is an endpoint. It collected 1,087 of these. It can now take 1087 different actions inside HubSpot.

It can now it can write its own make workflows, its own Zapier workflows, but it doesn't need to use make or Zapier or anything. It's gonna run off of your computer. Your computer is the engine that will run it. And the and this Hermes agent will be the will be the driver. All right, so what did it say? So one you requested up to 128,000 tokens, but can only afford 74,448. Okay, so this here, I wonder if I just

burned through all of my open router tokens because I only put a small amount of credits on this thing. So let's see, I just put no, but I did I burn through the $30 credit? I did.

Okay, we gotta delete that APK again. Hey, I burned through those tokens already on all the training, so can you increase the limit on that key to save $60? So I have another $30.

So now you might say, holy crap, Adam, $30? Like we've been on here for 34 minutes. Okay, what do you think it would cost to get somebody to read the API documentation? If you said, hey, I just want you to read this, not understand it, not like not document it into your tool set, right?

I just want you to read the instruction manuals for the entire Milwaukee tool set that is sold in a store today. I don't want you to add the tools in your toolbox. I don't want you to know when and where to use them. I just want you to read the instruction manuals and understand the 1,087 different tools that exist in the Milwaukee Tool Set Guide.

What would that cost you?

Yeah. It's a lot I mean it's more it's more about now it costs more of the time. That's it's you're doing it in record time instead of doing it spending it it just takes a lot of time and you're doing it really fast. That's why you're you're paying a little bit of money. But it's not you know relative speaking is not that a lot of not a lot of money 'cause it's it's it's more it's it's instead of a time thing, it's more of a a a money thing, right? It's it's more of it's more important you you don't spend a lot of time. That's an important thing.

Yeah, it's like you spent seconds doing this. Okay, so I'm gonna say now did it give me an update? Yes it did. Okay, so

Okay. Okay. Actually, we're gonna go say all right. So you have got some credits at this point, and you have done some documentation on HubSquat and Zuper. So a very common thing that happens in the roofing business is we want to see if a customer has signed a quote and what the quote amount was for.

So that we can celebrate the sale. And it would be cool if you could set up a cron that would look at my HubSpot and Zuper to validate that a quote was signed. And then every three hours let me know which quotes were signed and send me a celebration message in Telegram.

Now I'm running a agent harness out of credits is like running a diesel engine out of gas. So I predict there might be a bit of a hiccup here, but we'll see. Nope. Sometimes when you run out of tokens, it'll be like you have to start a new session, da da da da da. But

Yeah, yeah.

And so what it's gonna do now is it's going so it's figuring out like what did he mean by that, right? So it says, love this celebrating wins is great use of a cron. Let me first poke the super quotes API and learn the actual data shape, what signed looks like in the field. What fields are useful for the celebration message, right? And then it says, okay, slash estimate is the right path. Let me look at the data shape and find out what sign means.

And so this would normally be you clicking around inside make or zapier or n8n or gumloop or anyone lindy or any of these AI things, and you'd be create create a node and then create a branch, and then if signed, you know, if the signed date is in the future, and then you know, c then make sure you get the signed contract amount and then put that into a s then another node for like Slack or Telegram, send message, and then you drag the little JSON piece over to say.

This is a contract signed amount. So you write contract signed colon and then you drag the little H, the little J something. You have to do any of that.

You just talk to it and then it's gonna make the thing and you and then it's gonna make it repeatable and it does all the coding by itself.

It's like I'm expecting a room full of roofers to give me an applause. But it's like now we can see here, it's like, okay, let me check the docs for actual status enum and also check whether a separate proposal object in the dock. This is a brand new Zuper account, so they might not actually have any sign. That's kind of why I chose this one, was because it was a client with a that's just onboarding to Zuper. So we may not get this workflow to actually work, but it's not because it couldn't do it, it's just that like.

There is no signed quotes inside of this brand new, brand new account that is just in the process of rolling out. But I'm I'm hoping that we're kind of catching the drift of what's possible here.

Yeah, yeah.

So when it finishes trying to build this, I'll ask it to do something else so that people can get a little bit of like a confetti pop. We built something that worked.

And then I there's other exam I'm sure there's other other use cases too that you can they're gonna have it that you have it done before and maybe you can say this is the things they could do now.

But you can go ahead.

Yeah, so like we could give it the we give it our like Google has like a credentials file, right? And so you can upload your Google like Zoop, Google is slightly more complicated and secure than like just an API key. It's about a five-minute process. But then you get essentially the same thing, you get your key, and then it authorizes this machine to like sign in and use Google. So then you could say.

Anytime a signed quote comes in, I want you to write the customer a thank you message, but I want it to be written, you know, specifically about something that was noted in the quote. Like, did they pick an upgrade? So, like, if there were selected upgrades, I want you to thank from them for the upgrades. I want you to mention the roofing material that they signed off on. I want you to like, and instead of you writing a dynamic proposal or anything like that, it would just fire itself up.

And it probably wouldn't use Opus. You would say, hey, for when you write that email, I want you to use Gemini 2.5 flash. So that's kind of the cool thing about this, is that you can give it multiple API keys. So you don't have to use the 200 IQ super genius brain to write the email, right? It would just use something super cheap, something that's like 20 cents per million tokens. And then the execution around it might be running on Opus.

But that one step of writing the email only cost you a fraction of a penny.

And so the whole idea here is that we're trying to, we're trying to get AI to work for us. And so it's asking me to give it certain permissions. So this is something you're gonna see quite a bit when you first install it, is it's gonna be asking you to like let certain tools use your machine.

Still have to say yes a bunch of times. But okay, state is seated. What estimate 181 is recorded as already celebrated. Future runs will only fire when a new signed estimate appears. Now registering the cron in no agent mode, script only, no tokens, fires every three hours. So basically, now because it wrote a Python script, this is exactly what I was talking about. It wrote a Python script for how to do this. So it uses your computer's CPU, your computer's internet to be like a make.com. So you don't pay for

Make.com credits or N8N tokens or or even AI tokens, no large language. It's just gonna run a Python script, took like a little mini software on your computer that whenever the clock ticks over another three hours, it just runs this Python script like a little software and inquires with the thing. And now here we go. It's gonna say, okay, whenever we get this, it see, we didn't have to write this. It's gonna detect how it was signed in priority, real e-signature, sample celebration output.

And it's gonna go Albus Severus, dub out this on this day. And it's gonna say, here's the this is the estimate. So this has just been built. We didn't have to open make.com and figure out nodes and actions and branch criteria and loops and and wet webhook post actions. We have to figure any of this out.

Amazing. Awesome. Awesome stuff. Can you can you see that in HubSpot?

I could see this in Zuper, because this is in his in his Zuper, but I can I I think I have access to his hubspot. Let me quickly triple check here. That's the crazy thing is I don't even have access to everybody's hub spots, but I think I have access to this one.

Super, okay, gotcha, gotcha, yeah. All right.

And this guy was Marcus.

Where was that guy's name?

Yeah. No abus, abus, yeah, several yeah, yeah. Yeah.

Albus, that was it. Albus. I thought it was Marcus. There we go.

It's not appearing.

Almost.

Okay, well let's let's do something in HubSpot directly because maybe he's just a maybe that's an employee just testing inside of her. I want you to find me the most recent three contacts added to HubSpot and tell me the lead source that is in the the lead source box. And then I also want to know what the original traffic source was and tell me

Yeah, yeah.

What the response time was.

Фром ферст імел орст кол ноти о контакт.

So now that's this is not a skill yet. This is me just essentially inquiring into my CRM. So it's gonna write a script. This is a very expensive way to like find out the full first three contacts. But then you could say, okay, well, this is something I do all the time. So every day I want an audit of what new contacts were added to my system, what the lead source was, what the response time was by text and call and email and or whatever your CRM's tracking.

And I want you to email that window too small. there we go.

Want you to email that to the marketing guy to let him know what's going on. Okay, so there we go. you know what? That guy I was probably looking up a contact, but it's probably a company that's on the Zuber because this is commercial roofing. So lead source is blank on two of the three with one value being public bid. Response time is fuzzy here, created by inbound email. So the email's time second is four seconds before because.

Got it, got it.

So basically, the guy emailed the company, and because they emailed him, that's what created the contact in their HubSpot. The email is logged, it's reply, it was nearly silent. So a lot of this is coming in from their emails. By response time, you mean contact creation of first outbound email from your team. None of these three have a measure of response time yet. Okay, great. So we could say like how many deals were signed in the month of May so far. So these would be.

Deals that have a contract sign date or a closed one state or a deal stage that is like sold. I guess I want you to look through HubSpot and get clarity on what deal pipelines we have, what the deal stages are. And then I want you to see if you can start to see what deals are in what stages. You could use GraphQL if that's cheaper.

Or you could just do API calls, whatever you think. But HubSpot API does have GraphQL on your scopes. That was a bit technical, but in HubSpot, you also have this thing called QL query language. So you hear things like SQL or GraphQL. GraphQL is a thing that Facebook made. But essentially, a cheaper way to analyze huge amounts of data is with something like SQL.

And so an advantage to HubSwatts API is the ability to pull a large volume of data without having to write 30 API calls. It can just write one query in a script. And so I kind of threw at that bone to say, like, hey, let's not burn $15 on this discovery, right?

What what are the okay, so you're doing this right now. What the typical ones? This is kind of you're demonstrating the actual things they can do. Now, what is something that that is a business a good business use case that you that you that you've seen in, you know, home services roofing that you work with that can help them right away? Like like this is something, okay, that would help your like your business use case.

Yeah. So once it has a so what we're doing is we're training it to understand what exists inside this CRM, right? So once it knows what exists inside the CRM, it would understand like what are like every every day it can look to see what date, what what jobs are scheduled for tomorrow on the calendar. Cause it can see the job schedule date, it can see the address, it can see all the values or properties that exist.

But one super common one is we connect to open weather. Okay. And so open weather is just a weather API. All right. So you can you can basically see, right, what is happening all the weather that's happening. So current weather data, hourly forecast four days. Okay. So if we want to look at the if we wanna look at the weather, we can see like we can give it a lat long. So every Google Maps address, right?

Every add, I guess every street address can relate to a latitude longitude. So you could say, hey, go check tomorrow's jobs, convert the address to latitude longitude, and go find out if we're gonna get rain there. And so it'll pull, it'll run this API call. It'll look to see what the weather looks like for tomorrow's jobs. And then at 5 a.m., it'll run the same trigger and it'll say, hmm, that looks like the weather's trending in a bad way. And keep in mind,

Every day it's storing this to its memory, right? It's seeing how the weather's changing over time. It's saying, they were calling for rain three days ago. Now it's actually raining. But you can say, you can tell it. If the weather says that it's going to be a greater than 50% chance of rain and more than a quarter inch of rain, that's our threshold. Like remember, I said the tacit knowledge, right? So when when I ran a roofing company, when I ran a roofing company.

I would all we would always use AccuWeather, right? And this would be our process, right? We would say, okay, what's the weather look like? So we click on the map, right? And we're gonna get all these stupid ads. And we would zoom in, right? And that you'd be like, okay, this is where we are, and you'd hit the little play button and you'd see where these like little green blobs were, but you'd be like more zoomed in, right? And you'd be trying to see, are these are these green blobs looking like they're gonna come over?

The city over our jobs, stuff like that. And you'd look at all of this information and you'd look and see, like, okay, like this morning, today it rained until 1 PM and it was a 56% chance of rain. And when you look at that, and again, this is all extractable by the API, you're saying, okay, what was the chances of rain? 58% chance of rain and four millimeters. Not a ton of, not a ton, 0.4 millimeters, not a ton of rain.

So this doesn't mat meet the threshold, right? This would not have flagged. But when it extracts this, it would say, Hey, it looks like it's gonna rain tomorrow till 1 p.m. You could say, if it's gonna rain heavy until one, I want you to reschedule any replacement jobs to tomorrow. I want you to text the homeowners telling them, I want you to text the crews and tell them we're rescheduling till tomorrow. But I want you to get repairs that we have scheduled.

And schedule those repairs after 1 p.m. Cause we could at least get a couple repairs in today. And it would move all the stuff around inside the CRM for you and it would contact the customers. This is super common because then all you're doing is ch connecting the just call API, right? Which is another very common API. And you'd be able to say say, okay, well, I want you to be able to send an SMS. Check the reply.

Right? Send messages to a group. Get a group message. we just hired a new contractor. Add this member to the group. Right? List all threads from today. Did anybody get did like get check the reply? Did the customer complain we rescheduled the job? Okay, well, then go back to the open weather API and find out when is the weather going to be good. Then go into HubSpot and adjust their install date. Do you see what I'm saying?

Yeah, yeah. And let's let's kind of break it down to make sure you guys understand the value of this because it's it's really amazing. This is more more valuable when you see the more value for the bigger bigger companies. So if they're if they're doing let's say three hundred roofs a month, that's about three roofs a a a a a day, right? Or something like that. Ten roofs a day, I'm sorry. That you see how

Yeah.

Ye that would be one before that would be one full time person doing that, right? It would be one full time or maybe two. How how how would that work before this existed? How how would

Because there's a yeah, that that company has three production managers and they do about that many roofs and that there's literally a full time guy who's paid a six figure job just doing that one job.

And this would cost three bucks a day.

You see how that so you see how how that it and and it there's a lot of fear behind like it's gonna yeah, it is is it is gonna cause some some job loss for sure for some people in the company, in every industry, right? But you see how the company, if you if you actually set up the correctly, you're that person maybe is one person doing full time, they can maybe do something else. I mean, that really depends on the owner. I I I I don't like firing people.

But that that guy is running this agent. That company hired a guy to continue what this, like this roof claw, was their start into the AI world. So the guy who's doing that job, his job is working with this thing as a co-pilot, basically being a human in the loop on his decision making. They're working together and it's becoming more autonomous, right? Then they also hired a guy. So now they've created this is created a job.

Right now, what this has also done has made their scheduling more efficient and it's freed this guy up to help with other things. And the goal is they're opening up another branch. They have three locations, they're gonna open a fourth. Well, now they can promote from within, and instead of having to promote someone else into the position of this job, which was unfulfilling, like how rewarding is it to wake up every day and

Play Tetris with the with the calendar because Mother Nature decided to throw some wind your way. Like, how fulfilling is that job? Clicky clacky, tick, you know, toggle tacks going between the weather window and the job calendar and the crew's phone numbers and the customers and CRM phone numbers in the CRM, and then back to the calendar, and then keeping in your head that this crew is going to be working on these days, but not these days, and then Tuesday's a holiday, and that you promised Mr. Smith that he would get an install on Friday because he's a firefighter and he works nights.

And Friday's the only day you can install it. Like all that shit that you got to keep in your head. How fulfilling is that job? Not at all. That job sucks. You want to talk to customers and empathize with them. It just takes all the not human jobs away and gives t humans more time to be human and to serve more people because they don't have to, you don't have to allocate 40 to 60 percent of a person's job to clicking around in the CRM. When as you just saw, this thing just learned the entire

The entire API documentation for two softwares built skills and learned the company for 50 bucks in 38 minutes. And set up two cron jobs, which included celebrating a job and checking how many leads we got in the day before, what the response time was and what the lead source was. And we set that up in 55 minutes with some time to chat.

Yeah. this is something that just came to my mind. what do you think about something with you know, a big a big challenge is is is review generation, right? Review generation, getting reviews from from customers, asking for reviews. I mean something you can use you can do yeah, it's super simple now with this, right? How how would you build something with like with with with roof claw or something similar?

That's that's super simple.

so I very very simple. So I would say, okay, you know my HubSpot data. Every time a job goes from from install like installation in progress to install completed. So when the certificate of completion is signed or when the the ticket is closed or the deal stage moves, whatever your CRM does, because this works with Aculinks, Job Nimbus, RoofLink, Proline, whatever, right? But whatever the thing is, right? When that's done, I want you to go through.

And I want you to pull a picture from Company Cam that's from the from today, so the day the job was done. I want you to go through the emails and I want you to like pull the ticker of the tape, right? And then I want you to say, and then I want you to write a custom email that says, my gosh, the roof looks so good. I'm so glad the guys did a great job. Hey, I know we had a little bit of trouble there in the middle where we had to reschedule your job three times, but we really appreciate your patience and dealing with us through the weather.

And I'm just so thankful, and I really want you to enjoy your brand new roof. The HDZ in silver gray looks fantastic. And the gutter upgrade you got, it looked those new gutters make the house just look complete. You know, the guys worked really hard, and one thing that you know always helps us help them is to just show them that their job matters and they're appreciated. If you wouldn't find it to be too much trouble, if you could click this link and send a quick Google review, it'd be really appreciated. And then do something similar, much shorter by text, and then boom, just call API, boom, Google API.

Add a note to HubSpot saying we did it, and then follow up three days later if there's no Google review or no response to the text or the email. And that's where context layers and state control matter because it needs to know that the Google review happened so that it doesn't send a reminder when it doesn't need to.

Yeah, yeah, that that's amazing. That's amazing stuff. that that's

Which is way better than just like when the job's done, send them the link and say thanks for the opportunity to serve you. Can you send a Google View? Like this could actually read the seven emails that happened through the process and get an idea. It's like, what was this customer's experience like? So you'd say, like, I want you every every day, I want you to fire a cron that checks the jobs that moved into install complete. I want you to go to the company cam API. I want you to pull three pictures that are the most recent pictures, pick one that like looks like mostly the houses in the picture.

Yeah, yeah.

And then I want you to read the emails, find out what the customer's experience has been like, try and find a way to position it in a positive light, write it and write an email, and then take our Google review link, put it in there, and always and but always tell the story about how the guys work really hard and it's an unrewarding job. And we really like to show them that their job matters and that what they do mean something to people. And if you could do us a small favor to it's not too much trouble, leave a review on this link and let me know when you do. Then I want you to take that email.

summarize even shorter and do the same thing through text message and then wait for a Google review. And then every day, you know, or in 48 hours, I want you to fire another little check that checks to see if it's done. And it'll just build that for you. You don't have to go into make or zap your or NADAM and go into all the setup porn. You know, you just talk to your agent and tell tell it what are the closed loop and open loop processes that would give your customers the ideal experience that you want to be known for. And it will do the technical.

Yeah, yeah. Yeah.

Which puts me out of a job, but that's why I had to go start selling this.

Yeah. Yeah. And and that's my one of my I had a an event recently last week. And my my one of the big ideas is basically AI is is an operation strategy. It's not really it's it can be used for marketing, but a lot of people still think it's kind of a marketing angle to write copy and create videos and all that stuff. Genera generates things, but it's really, in my opinion, the the big big ticket, big big picture idea is that it's more an operations.

Yeah, 'cause it's it's generative ability, it generates things really quickly, so

strategy for your business. And what you're speaking about is operations, exactly, and making it more efficient.

Yeah. So again, for all my nerds in the room, like, yes, context layer, yes, state control, yes, cybersecurity, yes, memory, yes, yes, yes. I know drastic oversimplification, but for those of you who are inspired, right? I hope you take this and run with it. And if you're at the point where you're just like, yep, this is cool, but I want to set it up properly and I don't really want to step on 17 landmines to figure it out, like.

I'd be happy to talk to you about what setting one of these things like looks like for you, but yeah.

Yeah, awesome. Adam, anything else you want to cover?

No, that's pretty much everything. I just I hope I did a good enough job of giving some people some interesting an interesting venture into like what it looks like to work with AI today.

Yeah, for sure. What where do people find you?

they can go to roofing business partner.com or roofclaw.com. Roofclaws for that specific product.

Awesome. Yeah, and they'll have the links in the description. Adam, thanks man. Appreciate it. You too.

Hey man, no problem. Have a great day.