The Adam Sand Report - Thursday, September 3, 2026

Good evening, folks. It is another 9 PM livestream, just my way of having fun with anybody else who happens to follow along and join the chat. I like to do these livestreams at 9 o'clock because I know they'll suck. And that way, I don't have to worry about wasting any of your time. But if you are here, I like to work on agentic development workflows. I am building a product. I am investigating things, and I'm trying new things, and I'm learning every day. And so I like to connect with other people who are kind of on that same journey. And so if you have feedback or ideas, um, or ways for me to do better, I always appreciate the advice. But if you might happen to learn a few things as well, um, about my agentic development workflow, I would say I'm getting to the point of the borderline software factory. But it does mean that I still spend a fair bit of time up front planning, and then I dump everything into the software factory and let it run. Um, we are playing around with Fable 5. I'm waiting, anticipating Um, the, uh, the arrival of my Astra account. Hopefully we will get it soon. Um, and so yeah, we're, we are waiting to see. So maybe we'll just do one last quick check to see if we've got Astra yet, but I don't think so. So ChatGPT came out with their new Astra and it's apparently really ridiculous and everybody's raving about it. It's the only thing I see on Twitter right now. But I do not have it yet. And so, yeah, what we're working on today is a job file playbook prototype for a roofing company. And the biggest thing is that the biggest challenge has been training the model on how to analyze roofing jobs. Because there's a big difference between, hey, like having a system prompt with some AI widget that you strap on the side of a CRM that just says, hey, go read the notes and the calls and, you know, build a summary of this job with all the important parts. Don't make mistakes. Like, that is going to suck. And so if you guys are in the roofing space interested in, see, in, in agentic development, put a comment in the chat. I'd love to know who you are, where you're from. Um, we love to kind of engage with people in the chat and just like hear what you're up to, what you're working on. So, so throw something in there. But there's a big difference between that, like, hey AI, summarize this deal, don't make mistakes, And actually understanding like all the little intricacies that come in. And so the big project that I've been doing is building out the evals for how it analyzes a job file. And I think the easiest one is like to talk about silence, right? What happens when there's silence? So if there are, there are situations where something was never raised, like what are we doing with the satellite dish, right? And so then it has to either say, did I get nothing? Was there signal and we need to deal with it? Or was there some kind of negative thing? And then was it raised but it's unsettled? So a common case of this would be a customer said that they're choosing weathered wood shingles, but then they sent a last-minute text saying they'd like to go dual black. This was an issue that was raised, but the silence was from our side. Right? And so then there was nobody— there was no— nobody that actually settled that, that question, right? That would come up before you go to hand this job off to production. Um, half scene, right? So this is all of the different scenarios that it has to go through. And this is— I mean, we had to run like 477 tests on this, all these different kinds of situations. where what would, what would happen, because the, the goal of what we're building is trying to say, all right, if you're doing a job call playbook as a roofing business owner, right? So the sales team is gonna hand off to production. Well, sales is going to gather all the stuff, all the stuff that the customer knows, right? Because that's another piece of this is like, what is stuff that only a human would know? And what is stuff that we found in evidence, phone calls, text messages, emails? The, uh, the proposal and the photos. So like each photo has to have an AI analysis done on it, and then that photo needs to be turned into text so that it can do things like recommend gutter covers on a full gutter. But then when the sales confirms the stuff that the humans would know and the stuff that the machines would know and hands it off to production, there's a 15-minute production intake process that every production manager goes through. And our goal is to think in an order of magnitude of 10. So try trying to say, how do we make that 10 times faster? How do we do that in a minute and a half? And so, after essentially everything from here— wait, no, I'm on the wrong one. Everything from past the calendar. So, we went from the PM model, everything up from here has been a series of work to build out that those bedrock agents in AWS, right? So these are like Lambdas, and there's been a bunch of like work on Terraform, TypeScript, DynamoDB, all these things to like build these contracts of data. And then pulling a corpus of data from a client's account. So they're basically saying we pulled all of these jobs, right, from different clients. Everything that we pulled, we had to test the AI to say, can you extract these pieces of information But then when we put it in front of the— when we put it in front of the production manager, can they— and we had to build a prototype. So this is not the final. I gotta fire it up again. So when we put it in front of the production manager, is the production manager capable of reviewing the entire job file playbook and get the evidence that proves an answer on the question? without ever leaving the JobFall playbook. So typically in a roofing business, you're gonna have something that's handed over to you from sales, which is like, what's the material color? How much is the deal amount? What is the one thing that would make the customer mad? What is something that's out of the ordinary that was requested? What's their hot button that would make them a 5-star Google review? What is the color of the drip edge? Are we replacing the gutters? What are we doing with skylights? All of this stuff is information that's known either in the photos, in the proposal, in the measurements, in the communication between them and the customer, or the thing that happened inside the home. you have to record with some physical device and all has to come in and each one of those things needs its own analysis. And, but then if you, let's say you get all that stuff into the CRM and you answer all those questions and you put it like similar to just like a piece of paper that you hand off, well, then what happens is— Fire up the prototype. What happens is the production manager still has to go investigate. He still has to go and understand what existed in there. Our goal for this canvas that we're building is to make it so that that is no longer the case, that you do not have to leave the JAW Pub playbook to get your answers. I apologize. I was just wrapping this up. We are moving into the next linear issue. And so I shut all this down because we were closed. But, yeah, if you're in the room and you've been, and you've, you've, you've been sticking around this long, I'm curious about what you're looking for, what you'd like to hear about, what you're up to, who you are, where you're from. Hop in, because I'm gonna go to work, but I wanna give you an idea as to what I'm working on so that, you know, when I go quiet, maybe you guys have some questions. Alright. This, so this is not the final product, but this is how we used to demonstrate it. So this is again, this is an ugly prototype for what we're actually building. When we're when it's done, it goes into where is my right here? When we're done, it is in here. So it'll exist in this canvas where it's a canvas, it's a generative UI canvas that can change depending on what you're, depending on what you're doing. Right? And so it, whatever, whatever job you're doing, right? The, the canvas can look different. You can talk to it. It's like Tony Stark. It just generates a new user interface for whatever you're doing. So it's always kind of different. Let's go like that. So it's bigger. So it's like instead of the typical CRM, which is a menu on the left and everything else on the right, It's just you're doing— you're asking questions. You can be like, how's the business this month? And then you can make decisions, move these tiles around. You can resize them to be, you know, different sizes. But then as you make decisions, they go to the decision inbox over here. And then you, as you make decisions, you just— the agents go do the work. And so that's the premise of this whole thing. So it'll look more like this, but in this prototype, If we go to this here, the workflow, right? Again, this is ugly, but the workflow will be that these, these are present, these decisions are presented to you, and then you go, where did this— if this is the answer to the question, what is this customer's top button, the one thing that guarantees a 5-star review? They'll say, hey, financing payments, no surprises from the final number. Where did that come from? Right? And that came from comms. And so then you can see which call that came from and where it was said so that you can then confirm it and it becomes a locked decision. But the same thing would go for photos, right? So if it says, oh, there's this evidence for where do we put the material, where does it get dropped, that came from a photo, it's going to show the photo that that came from. So that's the whole idea, is that typically you have to then go investigate all these claims or understand why they're there. Instead, you just have the evidence pulled to the job file playbook from the Agentic review. So that's where we are and where we're kind of going. And it's the ugly one, or not that one. And then there's This one. Yeah, so this is a little bit more like what it would look like. So depending on the size you chose, so there's wide again, there's all different sizes of modules because in this it can be a different size module. So in the, in the prototype, we're just using something ugly to say, okay, what do I like, right? If I want to see the evidence, it pops up. If I want to go to this question, it goes here. If we look at like Mid-flow, you can see, okay, this is how many of the questions are currently confirmed, and these blue ones are what is curated by evidence. But while the AI is running it, it is confirming— that means that it's still analyzing to fill out these questions, right? And so the idea is that whether you have 6 sections or 9 sections, you don't have a different job file playbook for every trade. It dynamically produces the job file playbook depending on what trades are sold. So again, so this is like, again, the prototype. So whether you're using a— the small version, the medium version, and this is the whole concept of declarative generative user interface. Any questions? Because I'm going to go to work. All right, let's do it. So we're going to go into our new session. And we should have our new boot prompt. So my process for doing these things is I kick off with these boot prompts, and it all comes from Linear, from these issues that are generated. And we're using Matt Pocock skills. We can see that this is a grill section. So it's a grill me with questions. So you grill with questions. And if you remember, if you were here on my last one, our agentic development workflow essentially is Wayfinder. You basically— this is to break down a large subject, right? A large challenge, a large problem, a large task, something you're like, I don't know where to go. This helps you find your way. Then you can use the Grill Me with Docs where it will ask you questions to get shared understanding. You'll have an agentic decision register, and then you will essentially walk away with shared understanding of what it is you want to do. Then you can go to a prototype, then you can go /to-spec, and that'll build a product requirements document. Then you break that into issues, and manually you would just run a /implement or a /goal. But now we have it running into the software factory, so from— we go from issues to the planner agent. The planner agent writes a wave, basically how many non-conflicting lanes can we run in, where agents can be like software developers and work in a loop, right? And the orchestrator agent basically spawns and manages these ephemeral sessions that basically come alive, build something, loop through until it's done, die, go back to orchestrator agent. And the orchestrator agent just has to know how many independent waves can it run, or lanes in the wave can it run? Because the whole goal is to make it so that you're not— the whole goal is to make it so that you're not constantly building files that conflict with each other. You're not— they're not stepping on each other's toes. But then every time one of these things is considered to be done because it goes through its loop, then It goes back to the orchestrator agent who then fires up an independent verifier with another model. It verifies what was built, comes back to the orchestrator and says, okay, we're done. And if it's done, then the orchestrator agent merges. Once all of these are merged, it goes back to the planner agent and we go into the second wave, right, based on the issues that were logged in this spec. So where we are As we are still in this phase, right, prototyping and grilling, so that we can eventually go to spec to issues. And that's how we take the ugly prototype that I showed you and bring it over to Canvas. So once we do all of these grilling sessions, that's how we end up being able to just have a completely autonomous software factory. And it builds it for us, but we make sure we don't get slop out the other end because of the fact that we invest the time and the energy into actually doing it. So This here is me saying, okay, I have to then fire up this prompt for this right here. So I always like to review it. So, okay, what are we working on? Right? Actually, we'll go back to the pretty one. So what's already set? Okay. Yep. Grammar. Preconditions. So, yep, we're gonna do Wayfinder Grilling, Child of Plot. This is just my prefix and linear. If you guys don't know what linear is, Linear is like a project management tool, and so if I bring you over here, Linear, where is it? So it's just project management tools, so you can see your projects right here, and then in your projects you can have all of your issues, right? So if I go here to my issues under the platform engineering, Right, or foundation product, right? So I'm in this plat. These are all the— this is all the whole project. Basically, we just tackle these one at a time, and then these things go into review. Agents review them and close them. And so this is what keeps the software factory running, is keeping and filling it in. And so we have to go to plat 330. It's the one that we're working on right now. We don't have to go to it, but Basically, this is where it's managed, but that boot prompt is right here. And then anytime I'm making decisions or like, actually, I want to do this, actually, I want to do that, then it will put notes in here. So that's like, oh, that thing that you want that maybe isn't the best to work on right now, let's go put that onto the appropriate issue in Linear. I don't have to remember to go in here and do that. I seldom open this thing. I run it completely from Claude Code and Codex, and I can have Claude Code and Codex running on 3 computers because there's another computer over here. There's another computer behind me, and I can deploy and manage all this from Linear, and I can manage Linear from Claude. So I can create these issues, and then other agents can go work on them. All right, so these are the preconditions. Who checks? Okay. So what are we doing? The question from the ticket, the inputs per size information budget mapped to the full model. So, so the whale of the deal, that's like the what is the worst thing? An honest, empty, and unconfirmed suggestion in the packet only, or human words. Simulated review outcome, machine state, draft submitted, locked, accepted. Plant 327 is pending, failed the submit gate, disallowed sizes at production scale shapes. So it's essentially saying if Is there disallowed sizes? So when I'm on this canvas, if you're using generative UI, there are allowed shapes and sizes. Generative UI. So we look at essentially making it so that LLMs can generate their own user interface, but there's a degree declarative generative UI. So there's basically 3 levels to generative UI, which is static— you're very deterministic. So when you use ChatGPT to order a pizza, right, or DoorDash or Airbnb or something, or order an airline ticket, that, that is a very static user interface. It is very controlled, right? Then you have when you're just like, Hey Claude, build me an artifact, a report. Like when you saw that thing earlier, right? Where was it? Right here. When you saw this artifact that I made here explaining the evals that we built, this is a completely— like all the only thing that's consistent here is Mermaid diagrams, right? But that is more open-ended, right? So the agent literally draws it. In the middle, the agent fills it in, right? And so you provide it with atomic units of design, right? And then it can build from components. And so in this canvas, there are various sizes, right? Like, you can have a medium, wide, large, tall, but there's no small here, you'll notice, right? And on this one, it's just medium and tall. And then on this one, You can have a small, right? But you could also have a hero, right? And some will even have a takeover. Like if I go to Production Manager and I go schedule this week's jobs, this one, it has week, lanes, and month. Now when you're in the small size, right, you have hash— you have hover-over effects because the whole goal of generative UI is to minimize the visual text tax, right? They're constantly having to look at things, right, and read all the words. It's like they're— when you look at most, like, you look at our basically our backend, our backend is like every other tool you ever seen, right? It's menu on the left, stuff on the right, right? Tenants up here, right? Like, this is just, this is just the backend of it. But when I go here and we say, all right, well, who's— if we're going to have, um, if we're going to see, like, we see that this is rain right here on this day, we can see the weather pulls in here. But if we want to say, like, I want to have commissions also on this board for whatever reason, right? If you want to have commissions on this board, then you could bring this in, right? I don't know why you'd want commissions on here, but you could. But then we can also make this so that it will drag— it can resize things for you. Because then maybe you just want small weather, right? And small I should probably not allow, right? So that might be one of the sizes up here if that one's a much better one. Photos I can make small, right? And this one I can make medium. And then again, it's scrollable, but it's on here. But then this one goes to takeover and it takes up the whole screen, right? Then you can go to the month Crew lanes. There's no assigned jobs in here, but this is essentially how generative UI works. So what we're talking about on this latest issue, right, is we're working through the disallowed sizes at production scale shapes, right? So we're— the, the roofing companies that we partner with and buy, right, they run on average 10 jobs per day. Right. Right? So this needs to— this needs to be working at around 10 jobs per day. Some softwares make it easy for a roofer who does 1 job a day. It's less information. The UI is going to be different from a company that does 10 jobs per day, right? Because then you need space for 10 crews Monday. So this is what it's saying. So the Monday queue table, its own tall surface, surfacing before sizing, gratify, amend, or reject the candidate law. Okay, so this is about talking about how depending on what you're doing, right? I want this thing so that depending on what you're doing, it has some intelligence to say, well, this person's working on this and he currently has these cards on screen. So rather than loading up the default, the default, it might load up a certain size. To to make it so that the board displays a little more, a little more nicely, right? Like how would I say it? So that it's more logical the way that the board loads up, right? Sorry, I don't know why I brought that in. But anyways, so that's what that means. Delivery routing beside photos. Which module sees call and proposal evidence? Photos to photos is ruled R three. So. So this is saying which module receives call and proposal evidence. So when it's showing evidence, right, it might actually be an evidence module, right? Like, I might decide in this grilling session to make its own. So that's how this is going to go. This is how it's going to interact with me. So 1 question at a time, recommendation first, decision, recommendation, real downside, what the missing option D is. And then we're going to owner lock something in, and this is how it's going to record it. And that's going to keep it honest. And then we always do the cost because we keep track of the cost of the software factory because I'm trying to bring the costs down, right? It's like we were, we were running around $175 a PR and then $75 a PR. And so this is how we get it cheaper over time, right? So. Now, for the sake of making it entertaining for you guys, Everybody seems to like it better when they get to see me burn tokens, so we'll burn some tokens. I'll use Fable. All right, what else can we work on tonight? What else? What can be our side project while we're waiting for stuff? Muhammad, I will try linear. That's a great feature. Yes, it is an awesome feature having that for linear. I, uh, I love being able to orchestrate everything right from here. I mean, now that they have the browser window, it's nice. You can, you can open up Linear in here, but I will say I still like the Codex harness just a little bit more. But we're just waiting to see. Nope, still no Astra. Any updates? Nope, just waiting on the Astro to come in. So another thing I like to do is work with a coding and/or development partner. And so when I go in here, I have this Hermes agent that I work with. I'll say, at Markus. Slash new. Hey, we have fired up the lane F boot prompt for Flat330. Put that in there. By the way, something I've been working on, um, is that every time a new model comes out, I like to run some, like, basically tests and benchmarks and get to a place where I can have a skill called decode effort. And so this one here is now ready. We've been running tests on Fable 5.1 all day. And this allows you to— this allows you to— it's like one of my agents there, but this allows you to actually decode The effort. This is open source. You guys can have this. I'll post the link in the in the comments, in the chat here. And so what this does is allows you to run a say like a slash command of decode effort, and you can like pick like an issue out of linear, for example. And what it will do is it will go through a process, a skill of saying, hey, how do we How, how do we decode the effort of this task and make it easier for you to know which model and skill you should use? And so if you are ever sitting there wondering, like, oh, how hard is this task that I'm about to do? Because again, if you're following this process and you end up creating a bunch of issues, so then your linear is now full of a bunch of issues and you're like, well, this plat 218, Like, how hard is this issue? Well, then in your coding engine of choice, you can run this /decodeeffort and then your linear issue, and it will go look up that issue and it'll say, hey, this is how difficult it is. This is which model you should use, a Workhorse or a Frontier. This will work with Codex or Claude. It'll say, hey, this is how much you should— this is what you should use for Claude. This is what you should use for— there's also one for a Codex version as well. So I basically have 2. But then you can use this. It'll tell you what you should use. And then every time a new model comes out, I run this, this same process again. And so whenever I do get Astra, then I will kick this thing into gear and we will run a whole bunch of, uh, we will run a whole bunch of commits against it, then pull that in and say, okay, this is what you should do. So this is something that I'm doing just to put something out there. It's nice. I don't have any stars yet. If you want to give me a star, that's great. But it's something that I just have been using for myself personally, and it helps me avoid burning too many tokens unless I'm on a live stream and you guys just want to see me burn money. Okay, so what is it telling me we're gonna do? We're gonna go through these different cards, these questions. So these are the questions that it's gonna ask me, and these are the— my answers are gonna be named this. So you'll see that, like, it'll say G1, and these are the questions that I have to answer. So this is the part where it might— yeah, decode effort in GitHub. Decode effort in GitHub. If you don't see it, how would I make it easier for you? If you don't see it in the chat, Because you're on LinkedIn, then I will go and post it to LinkedIn right now as well. Updated, updated the Claude Code Decode effort scale for Fable 5.1. There, now it's on my LinkedIn. Okay, so this is the part where it gets boring because I have to actually pay attention to this stuff. I will take it to the agent so I can try and make it a little bit faster. So I'm just going to say copy this whole thing. Come over here. Say, okay, I am on the live stream tonight, so be careful. But here is the response from the boot prompt. I am trying to work a little faster so you can disseminate what these questions are a little more plainly. So I can share them with the people on the stream. Hey man, you no problem, Muhammad. All right, so it basically said this is a great— this is a grilling session. Yeah, how I'll work: post one question, I'll decode first, maybe missing D, then both. So Marcus, which again is a Hermes agent that is running, I believe, Groq 4.6. Yeah, it is. It basically just works like a little coding partner. So, decision, your candidate law verbatim from the September 3rd session before choosing a size. The composer decides whether a module should surface at all, plus the 6-step or healthy and silent, which role status, investigation or action lifecycle, exception state, board occupancy, footprint, record as intent times role times lifespan times exception state, board occupancy, surfacing and footprint, ratify, amend, or reject. Ratify this sentence verbatim as law. Ratify the 6 steps as the composer. So the composer is This down here. This is the composer. So this is the essentially the harness that sits on top of the canvas and is going to essentially have these decisions to make. Ratify the six steps as composer decision order for what it volunteers onto a board. So A one silence is a default, never a wall. Already binding through the spotlight through that blog, but it should be written into this lost text. So we read step one as the Label cannot be summoned. Well, healthy. Plain speak. Jarvis does not mention the suit's battery. So plain speak, it's saying this is it. Jarvis does not mention the suit's battery when it is full, but he answers the moment Tony asks. Healthy is proven, never assumed. The system stays on only when the state model reports it. No exception, no fail, no pending the 15-minute SLA. 24 split physics versus policy. Steps 1 and 6 are physics and become law. Real downside: law lives in the agent's judgment, not in deterministic code, so it cannot be unit tested the way it is for— this is a problem. So I'm actually So I'm going to come back to Marcus. Before I ratify that decision, I want to make sure that we are still in line with the Copilot Kit generative UI, A2UI concepts, where this is a declarative module, a component that is like a component generation process that aligns with that spec. I just wanna make sure we're not drifting because it's been a while since we've actually talked about this. But now that the prototype is coming into the composer state, I just wanna make sure we haven't deviated from this product plan. And again, for the other people that are in the room, would love to hear where you're from, what you're up to, what you're working on tonight. Everybody's always so quiet. I don't know why. Your finger's broken? I appreciate Muhammad. I appreciate you keeping me company tonight, buddy. Sometimes I like to listen to music while this is going on. That it quiets the silence of the rest of you guys not commenting. So this Marcus agent is running on this one of the two Mac Minis that are over here, but essentially it has like all my memories, all my principles. It's it kind of just like knows my preferences, whereas. Like in this case, Claude is acting a lot more like a product manager interviewing me to say like, hey, what is it that you want? Like here, let me give you some ideas as to like what the implications of those decisions might be. Let's gain a shared understanding before we go and develop a spec for this feature that you wanna actually put together after you saw a prototype. That's how Claude is acting right now, right? But something needs to say, Hey, you know, are we staying close to my bigger, larger vision? So this is not drifting. G1 is the 2-step. It's not new product. Okay, good. The last row is why A1 is not optional. Without healthy and silent becomes loaded bug. You already know why you're not— what you're not deciding tonight. Yeah, rectify A1 and A2. Genuine authenticity, placement, signing changes mind, go. Okay. Perfect. Now I'm just gonna go back and read that original comment that they gave. Okay. I feel like we should get Codex doing something just because I am getting those extra limits. What should we work on? Yeah, I don't know. I don't wanna— I don't wanna admit I'm actually enjoying just the focus that's going into this, as much as I don't want to bore you guys. See, by the way, this guy definitely want to follow him. I am interested in this skill repo that he was talking about. This guy apparently has jumped up on some stars lately, but he had a pretty good skill. That was— Show Me is a phenomenal skill. Yeah. Show Me helps the user understand the current topic visually with concise diagrams, code escapes, and HTML artifacts. How about breaking down actual log file data to see how AI scrapers, Claude and GPT, crawl sites versus normal users? I don't understand the question, Mohad. The actual log file data, see how AI scrapers Cross-sites versus normal users. Like, are you talking like a GEO, like SEO kind of project? Oh, you put in there brackets Abe. Maybe I'll just go to calling you Abe. Or how log, how log data shows whether AI search tools are actually parsing your business info or ignoring it completely. You know, one of, we had a hackathon here actually just last week and one of my team actually built something like that. They use this source, a data source called Data for SEO. But then I also saw today in a YouTube video about something that actually addressed that. I believe it was Colm Dean. His The AI data impulse. This allows you to actually see the logs because it'll give you these, these IP proxies. This would be a good pathway to go down. I don't really have much of a— but what my team did in their hackathon is they used data for SEO. Actually, if I go log in here Yeah, this is where he was working. I don't even know what he did. I guess I can't see the results anymore. What did he do? So basically they built a tool that goes and reverse, reverse analyzes your HubSpot data to say, okay, let's look at all your deals. Let's look at what keywords resulted in the contacts that are attached to those deals. This is how much revenue you made from this keyword. You should be spending more money on this keyword. You should be investing more time and effort into these keywords. That was something that they built the hackathon. If you ever come to our Tech Support Tuesday webinar, so if you go to roofingbusinesspartner.com and you go to courses, if you want to learn more about that, come to Tech Support Tuesday hosted by Gary and Matthew. They are 2 technical wizards that work here, but they would, they would probably know a fair bit about that. It just seems like they— like everybody and their uncle has a tool about the tool that does something like that right now. So it's not something I really wanna— not a rabbit hole I wanna go down. Yeah. Okay. So it ran that G1 ruling. So now what is this one? The whale precise. Again, I'll just to save time, I'm gonna rather than waiting to see if I have a question, I'm just gonna shoot it in. Again, make sure we stay principled, principled in our decisions here, and you keep me on track with the big vision. But here's the next question. All right, so the whale per size. So what I mean by the whale is that every job in a roofing business, there's like the whale of the deal, Right? So, you know, a person builds a garage on the side of their house, they tie the garage right into the house, but then you have a sloped roof on one side of the house and sloped roof on the garage. And because it's tied into the wall of the house, you basically end up with a dead valley running into the wall. So the water runs down, hits this dead valley, and it was fine when it was brand new build, but then now that they're replacing the roof, it's going to be disturbed. That's the whale of the deal. If you do nothing else If you get that part right, the rest of the roof probably won't leak. So the whale of the deal is something that you're going to pull out of the conversations. It's something you're going to pull out of the photo. And the whole idea is that if you're a production manager and you're trying to cut that 15 minutes down to 1.5 minutes, you have to be able to know what is the whale of the deal on this job without having to go digging and looking at everything and trying to figure it out. It has to be able to pull that whale of the deal. That's what it's talking about. So what forces it? Between playbook lock and PM confirmation, the whale field renders honest empty on every surface. The AI suggestion appears in exactly one place. The production intake review, the packet visibly labeled as unconfirmed AI suggestion. R11, the PM never edits the playbook. A human always owns the field. The ritual demo's own model agrees. The rundown computes the suggestion from the hot button and offset— wait a minute. The ritual demo model agrees. The rundown computes the suggestion from the hot button offset answers. Now, that's not necessarily true. So I'm gonna actually send that right over to Marcus right away. Before I even answer, one thing that I caught was it says the whale is taking from the hot button and what would cause a bad review. That isn't the only thing that computes the whale of the deal. The whale of the deal also comes from the photo analysis, right? Like we, we would annotate photos and The sales rep would identify, like, a dead valley or a chimney that needs a cricket or ground access that's a problem or potential pieces of the house to get damaged. There are things that could be pulled into the whale of the deal from the photos. One of the very original examples we used was a garage that was built and tied into the house, and the sloped roof of the garage runs right into against the house. And then that becomes a dead valley that leaks. That was one of the exact examples we took from Total Exteriors. So before I even send what you said, I just want to make sure I put that in there, because as I'm reading it, I see there being a potential conflict in that it's only taking from 2 fields, but there's a myriad of places that the whale of the deal would come from. And this is, if not the most critical field on the whole Jobfall playbook, is That whale of the deal. That's the thing that based on everything that it analyzed, it has to say, if we do nothing else but get this one thing right on this job, could be a site prep issue. It could be a material selection. It could be a long lead item. It could be a financing thing. It could be that the— that there's some kind of particular install detail that's really difficult, a material that's really tough to get. Alright. Like it could be anything and it has to rank what is the most important thing. I feel like this might have got missed. So before I even— before we even answer, I'm going to see what this says. Okay, so do not put the whale in the playbook. Don't show the eye of the guest there either. Okay, surface what it may show, what it may not. You show playbook every size and tell them if they need to act, lock, not a good extension, missing deal of the playbook. Looks nice. Okay, going back over here. Yes. Never appears in the play module, any size takeover included. Protected field has 2 honest states. First size budget for the protection action. The open action awaited with protection should route to the packet module, finding 6 constrained routes. Every size renders the field's current value only. It's the whale on Henderson is answered by the answer actually. Yep, that's so. This is basically saying you can ask what is the whale of the deal on a particular job directly from the composer. That's what that last bit is. Okay, the strongest attack is the PM Gladshaw showed suggestions. The intake moves faster in exactly one place kills it. The field speed promise the package is a second attack. Feel downside missing option D. Let hero take over host the post. Say action cable rejected, rejected. Never display with one exact confirm correct. All right, that's probably gonna say we're not deciding that here, but it's something I want to bring. So you're right, don't send the last page cable mixed to demo shortcut with the product original computer to guess what happened. I said that's a fixture, not, not the like the But it's not the whale. There was one right sentence we get to start the job. Say this composed from everything found in the human. Photos of first class dead valley cricket. Yeah, Raj type. Ready to go. Same class. So G2's home is still right. Say this instead. Okay. One thing I really wanna do is actually there's— there is one thing that I kinda wanna do. I want to— there's a particular website project that I kinda liked, and I, I saw something that was kinda cool from, from Nate Herk. I do wanna try this. Yeah, it was this. Have the same feel. And I explained all the elements of this. We've got the text, we've got the hero section. Those one website that he designed as an example without from, but instead sideways. And of course I used my Scrollcraft trying to find it. Yes, this. I love this. So this website had like 4 layers, right? Because you had like the mountains in the back, and then you had a second layer of mountains. So watch as he scrolls. There's like this— see how there's like 1, 2, there's 4 layers? The back mountains, the closer mountains, the screen, and then the And then this. And I thought when I was looking at this, how cool would it be to have rows of houses, right? Like houses in the far distance. Like, that would be cool to kind of have houses. And so I wanted to find a way to replicate this. And in this video, he talked about how he did it. I thought I joined this school already because it's free, right? Curious if he has his examples. You could take this, take this to Codex. Actually, we'll go to here. I'll go to Do you remember the conversation that we— well, conversation we had about the website redesign? I watched this video the other day that I thought was really cool. And it showed this website of that had like four layers of mountains. All right, so it was like three layers of mountains with a computer screen in the middle, and I just I thought that'd be a cool website content or concept to build. He built he built a skill that apparently has it inside. I'm currently on a live stream. I thought it'd be neat to show. Maybe use some of our resets on Codex while we're waiting for Astra to build out a concept for it. Excuse me. Maybe we'll work on that tonight. Machine state and simulated review. Okay, so the simulated sales manager review is basically after the salesperson fills out the drop-off playbook, sales managers typically have to review it. The problem is 80% of the time they review it, it's fine. And when they're— when it's not fine, 80% of the time it's just not complete. Right? Only 20% of the time is it done wrong, and only 20% of the times that it's done wrong is it something that actually requires them to look at it. So that's one of the agents that's in there. So again, we won't, uh, wait. We'll just send the question in right away. Okay, that is sent. Here is G3. G3. But then when I don't read it first, then I always end up having a secondary question. It's happened twice in a row already. Okay, machine draft locked. Can reject back to draft? Yes. Public session left render prep still running explicitly rather than showing a thin line draft. Okay, yes. Rulings done for failed candidate loss 6.2 were pending. Where is a pattern? Machine stays 1 declared chip on every size, blinds included. Prepping is distinct visible state, never a thin craft. A guy in a glass wears a thin running pattern. Yeah. A cell count interview sizes show review outcome has 3 values: rejected, budget, versus glances carry 1 rejection badge with a printed count. Percent reasons are on hover, focus, and tap. Yep. Flag delta is the same visual species as staged new evidence. So visual species is kind of just saying like atomic design, right? So it's saying it's— we're going to create something new, but it's going to follow the current design room because there's a whole design room in this repo of like, what are the atomic units of design? So everything has the same— everything in the generative canvas uses a similar design language. And we're even tightening this up as we go by like choosing the red dashed line to show like that this is weather. That was something like, that's like, okay, how else are we going to use this red dashed line In other modules. Right? Yeah, I like the missing D. Basically, what it's saying here is that in that prototype, as we were showing you before. Oh, is it gonna— ah, it fired. It's turned off again. Yeah, so basically it's saying on the small size, if the playbook was— the job file playbook is rejected, it doesn't have enough room to show all the stuff, right? So then how would it show, right? It would have to have a different effect here. That's what the, the D Session is saying right here. We'll see what Marcus says. I kind of agree with what is said about the recommendation that wasn't mentioned. It's interesting, it's not asking me, like, it's not giving me 3 answers, it's just giving me a D option, which is kind of weird. But I do kind of agree that the small size would need a different kind of provenance minted or some kind of different chip. So pay particular attention to that. All right, what's Claudia working on? Yeah. I agree with this. Now, instead of the 4 layers— back mountain, middle mountain, laptop screen, and front layer of trees— I would want to make it rows of houses. So it'd look like houses with a relatively probably closer, um, aspect ratio, I think is the language I'm looking for. So it'd be like Row of houses that are close, row of houses that are medium, and then maybe a cityscape in the back with that like variable parallax effect. That was kind of what I was thinking when I saw it. I thought it'd be kind of a cool project to work on from that forest site example. So we get which size is submit or lock. What is our yellow? Why does it not submit surface? This confirms the state-changing action belongs on the action-capable surface. This is not real weather medium. Interesting. One thing I found interesting is it says that medium is not a surface that allows for submit. I find that to be a little bit Off, why wouldn't you be able to submit from the medium? Why would you have to alternately change the module size just to hit submit? Am I getting something wrong here in my interpretation of what it said? Yeah, because that doesn't make any sense. Why would you have to change from the medium? This rule does not rule whether media may submit the form. It seems like they're going to get a counter notice about pending precondition. See, I'm reading too fast. Yeah, see, submit lives in every action. Failed gate submit. Ratify 6.2. Small, medium, large, tall, hero, and takeover. Okay. I think it's gonna— I know it's gonna be Yeah. Colony was— this open claw is going slow. It's going kind of slow. So, for those of you just hopping in, we are working on a couple different projects. One I'm kind of half-ass working on, but we are making some decisions on a roofing job call playbook automation agent within our Foundation Canvas product. And so we are just banging through kind of some last-minute decision criteria, some really just really important product design decisions. And I'm kind of working, trying to work a little quicker by submitting the questions to my Hermes agent for a second look so I can go faster. So you guys aren't bored just watching me, but it does mean you're a little bit of just like watching me work today. I like to do these things with you guys around. Makes it fun, keeps me accountable to working, and I sometimes tend to connect with some cool people. There are 3 sweep clear roll 7 slides in 1 sheet. Local 4, so that was a— that was a customer. 36 sections, 33 items. Here, 103 log, 9-section stress mode harness. We're going to send this to— we're going to send this to the Hermes agent. Is— have not, not read it yet. Yet may come back with before. Take your recommendation. Oh, we should check Codex again, see if we got our— see if we got Astra update yet. Nope, don't think so. No, Air 15, go to scheduling. Okay, this ticket grants no production size. Disallow sizes at production scale shapes. What grants the size? This ticket grants an over every row on the table, except clear for the oak 4 shape only. The playbook declares it allows sizes starting empty and earns each one, each one the way the ticket was earned. Explicit interest after design room proof. 2 production scale shapes declares data S1 is a 3 trade deal. Riffin plus gutters. Yeah. S1 is plus per orb overlay is declared maximum error 14. What each size must prove to us are 3 assertions: fixed frame, 0 nested scroll, 0 clip controls, or yes. Okay, nothing is restricted today. Missing option D, crash specifications is enforced. Okay. Real downset, the module cannot be placed on any production board till the design room slice proves S1. The spec must request the proof. First implementation slides. Yep, that makes sense. Now, what were the original decisions we had to make? We had— oh, we had— we are on this last slide, so we still have to make 5 more decisions. But this is what makes a good product. It's worth the work. It's worth the work. Because what comes after? Yeah, so if we just have 3, 330, 324 needs my CTO's approval, 331, and then these ones here, these are resolved. That we just did. So we are this one and this one and CTO review away from going into spec. OK. Man, there's like three of you that are persistently here. I don't know what you guys are doing, but appreciate you guys hanging out. I'm not even playing music this week. This already carries a Monday job playbook declared decent modules through the lounge sizes. Yeah, that's the old version. Monday queue table, the rituals demo fifty fifty dozen crew fifty to eighty job scale out numbers. Its columns are fixed fields. Yeah, its own module type. Decide, is the Monday queue its own module type? I believe it is. All right, everybody. Well, I'm starting to get tired, but I gotta get through these. I gotta get through these 10 questions. I gotta get through these 10 questions. It's probably gonna take me another hour. But then that'll be from 5 in the morning till midnight. That's a long day at the desk. Of course, we had dinner and walked the dogs twice, but I was really hoping I'd see Astro come in tonight. I imagine we'll eventually develop a module that is a list. I, I definitely think people are gonna wanna see a list of things, and that's a module we just haven't built yet, but we will have to build an atomic list module that could be used across all matters of things, but that's a later problem. Takeover reachability beyond glance resume. A hero's eligibility or owner lock requirement. Stage anchor, the placement holding the one stage capable footprint precedes takeover hero. In composition, may place a module at takeover with no act involved. May place a module at takeover with no act involved. Spotlight may present takeover only where the module has earned it. Size revolver skips the user's override on a stage lend and applies the destination for resume when takeover is not being earned or placed. Composer image surface or not, the clearance is 3x, never 1 bit. Takeover is earned by proof, same as every other size now. A rep working at large who wants the combined view has 2 steps. Ask the composer to resize the monitor. There is no button on the tile. Your lock accepts that trade-off for reading the monitor and choosing its own size. So. There is a future design for the canvas that is not considered here and may not need to be, but I do eventually want to add the feature to be able to move the tiles around because sometimes they arrange in a disorganized way. In HubSpot reports, for example, you can drag corners of modules out to resize them to set sizes, and you can move like tiles around almost. I'd like to have that feature later. I don't think now's the time to litigate that, but I just don't want— I just wanted that to be said before I give you the next question so that you understand Something that is not existing yet in the frame of our conversation or vision, but something that I definitely want to implement at some point. So I guess what I would like is some kind of collection of resources related to this plan that I could put somewhere to run a Wayfinder session and then Grill Me with Docs to then take this to design. That's something that's on my to-do list. Something I'm not really super passionate about right now, but I thought maybe I could get into it tonight. Yeah. Yep. Can you park that HubSpot resize and drag move and snap feature for later somewhere in the year so that we can bring this up at some point? It would go on the project as an issue in the backlog. It's something I haven't brought up before, but it needs to be put up there. You can tell I'm getting tired, but I don't want to be lazy. But come on. Oh, we got 3 more. What module receives column proposal evidence? Composer phrases and agent inputs and agent agenda entry. Man, I really don't even want to do 3 more. I'm just wiped. Do I survive on caffeine, nicotine, and protein? Controversies, con proposal evidence. What evidence in the photo module are you ruling Amendment 5 preview and routing exceptions? Yes, I'll call yes. In the prototype, it was kind of minting its own pain, but I knew that we would get here and have to deeply consider this. So here we are. What do you think we should do? I don't know if I want to have a separate evidence pane. I feel like there's already modules that have this, but it would be weird if it's constantly snapping new modules in and out, which does present somewhat of an argument for having an evidence pane. Some kind of evidence module. But I mean, really, what is it going to be? It's either going to be like from the photos module. It'll either be a blurb of text from an SMS, email, or call transcript, or an in-home recording transcript, or it's going to be from the inspection checklist, or it's going to be from the proposal. The hardest one would be the proposal because the proposal is going to be one kind of large, almost like web document or PDF type thing. So in some ways, Most of it is going to be text or photos. So I don't know, this is a tough one to consider. I'll give you the question and you can help me work through it. Oh, ran out of, ran out of credits for console. One second. Sorry, I gotta pay for some more XAI credits, my Grok 4.6. So I'm not gonna put billing and credit cards up on the screen. We gotta move this over here. I think I might call it quits here, guys. It's actually great having you guys here. All you're all quiet. Would love to chat with you, Muhammad, Abe. It was good chatting with you. Have a wonderful night, everybody. Thank you.