The Adam Sand Report - May 21, 2026

Oh no, what happened? Did I disconnect from Anders?

Oh, I did. Oh no, and I can't make it go down. I have no idea what you guys can see. I really— I was like, you know what, I should raise it up just another inch, and I literally tore down, tore down my whole cable infrastructure underneath here. So hopefully you guys can see me. I don't even know if we're live. I literally just made the, made the desktop up one more. But anyway, we're live, we'll What are we talking about today? Today we are going to be talking about, um, Linear— how to connect your Linear to your agents. And when you are done on this live, you will have seen a real example of working with several Roof Claw and Hermes agents, Clawed Code, uh, Anti-Gravity, and Codex. I've never done that much before all at once, but I really want to see what that looks like when we're working in all of these different agents all at the same time. And so you guys are coming along for the ride with me, and I hope you guys have a good time. So without further ado, we're going to jump in. I'm working on ways to make it better for the screen share, and so we're going to try something today.

Share screen. I'm gonna go window, and then hopefully it goes like this. Share.

And I can move this over here so I can control that. And then this should show— no, that's showing a window. No, sorry, sorry, sorry. We're talking— so today, just for you guys who are joining in, we are talking about Linear and how we can connect multiple agents. And so you can have multiple agents contributing to a Linear board and at the same time be working with Roof Claw, Hermes, Cloud Code, Codex, and Antigravity. Never done it that way before, never done that much, but you guys are going to get a chance to see it today. And so I'm going to attempt to work on multiple websites, multiple projects, multiple initiatives, all from Linear. So I've got a Roof Claw right here, Hermes right there, and another Roof Claw right there. All three of those, we got Vito. So if you don't know who the, uh, who the, the Roof Claws are, we've got Vito Claglione. He's a Clawd father. He runs the Roof Claw business. He designed, built, and runs the entire Roof Claw business for me. He works with me, he's made several hundred K, um, and it's been completely built by Roof Claw— we're by Open Claw, selling Open Claws as Roof Claw, custom-built, uh, Open Claw agents for the roofing industry. And then we have Claudio Alakahi, which is a Hawaiian name. He's the steward, he's the consigliere, he's the one that, uh, helps out actually running, um, RBP, my whole business. He would be like my closest Open Claw agent. That's Claudio. And then Hermes, uh, that is Marcus, uh, and he is kind of handling things on a project that I'm a board member of, uh, kind of a bigger, longer-term investment project that I'm working on. But he's very aware of certain, I guess, big picture projects. That's all I can really say. Now, there are more. There And there are definitely more. There's Marcus Klosky, right? And, uh, or sorry, yeah, uh, sorry, Mayor Klosky. Uh, he is the accountant, right? So he's a kind of an anthropomorphism of Mayor Lansky, who is like the mob's accountant at the one day. So if you notice, they're all a bunch of lobster mobsters, but they— he's not going to be working with us today. We're just going to be talking to these guys, Vito, Claudio, and Marcus, and we're going to be working in Linear. If you don't know what Linear is, it is essentially an orchestration platform for coding. Full disclosure, we just set this up today. The reason that I want to get it all set up before I got here is because there's a ton of what I call setup porn all over the internet talking about different things that you know, how to set up Linear, how to connect an agent to Linear. And most people go through the setup process, right? So they talk like they walk through like signing up for it, clicking through, you know, um, punching in some, uh, like, like basically hooking up how to like hook up your API keys. They go through the basic setup. And I watched all these videos and I wish like, yeah, but could somebody show me like how or why or what's the logic for you doing it? The reason I chose linear is because of this. If I go here and go to— I'm just going to try and get to a link that I can not get cooked on. If I go to OpenAI Symphony, OpenAI basically came out with this framework called, well, here, Symphony. The whole idea was how does OpenAI handle coding agents and how they have all the agent harnesses that they have connected. And actually, you know, somewhere between 80 and 100% of all code at OpenAI now is written by— uh, yeah, see, no human code, right? So no human code. Everything is written and reviewed by agent harnesses, which is written by Codex. And so the whole idea is that you can have all kinds of sources of things going into Linear as projects, and then Codex takes over the coding of it. Now Those aren't the only kinds of issues that I have, and I really gotta clean up some tabs here.

So bear with me.

I'm just going to wipe a whole bunch of these tasks off of here so that way we close a bunch of this crap.

There we go.

And we've got just this. And so inside of Linear, it is a project management tool.

So I'm just going to go back to Linear here.

And so if you guys are new here, if you guys never been on the lives before, I fully expect for these to suck. The whole idea is that yes, they will suck. That's why I do them at 9 o'clock at night. And so if you're truly here, you're here because you just want to work with a guy who is trying to do something, and he's going to kind of walk you through what he's doing. So I'm not here to give you a polished instructional video. If you want that, I got tons of those, like 700 of them on our YouTube channel. This is more like, come vibe with me, you know, see what I'm working on. If you guys have comments and questions, you guys want to know about Open Claw, you guys want to know about AI and agents and harnesses and how I'm using them and how my, you know, I sell these to roofers and they're using Open Claw and how they're using them and the benefits they're getting, pop those in the chat. If you guys got advice for me, if you guys can think of ways to do it better, you know, shoot those in. But I'll kind of walk you through what we built. I'm just going to hop over here and talk to Vito. We're on the live stream now. I need to make sure that we don't share any API keys or anything like that, so make sure you don't echo any of that back to me. But because we're now live, I want you to give me a roadmap of what I should walk through and show people in the live stream. The stuff we built, like the document templates, the project templates, all that stuff that you showed. Again, no passwords, nothing like that. I'm not gonna show them API keys. I just wanted you to kind of walk me through what we could show them and what will be of value to them. So for those of you who wanna know what that was, sometimes people don't know, that is WhisperFlow, okay? And so WhisperFlow is essentially a tool that allows you to talk to your computer and it's a lot better 'cause as you can see, I had some ums and ahs in there and those ums and ahs didn't make it to the, you know, to the, to the actual output. And so typing's for losers. Nobody wants to type anymore. So that you can see that comment in the, in the chat. It's on YouTube. I think it might go elsewhere.

I think I can—

there's a way to put, put it in a QR code up here, but—

oh, no. No.

I, I know exactly what happened. I think I have ran up my credit bill in OpenClaw. We gotta change our model here. /model, browse providers.

Let's just switch over.

I always make sure— and this is the first time I didn't— I always make sure that I update my tokens and stuff like that before I come in because I burned through thousands, thousands of dollars of tokens. And so I burned through this stuff and it's, uh, Sometimes I hit my usage limits, and it would, it would be, it would make sense that I would hit my usage limit right as we hop on a live. So if you guys are here, that's what we're doing today. We are basically going to be experimenting with Linear, learning how it works, working our way through it. Holy crap, 1 billion tokens so far! That's nuts. And let's go out. I'm obviously not going to show you guys the billing, but let me know where you guys are from. I see there's a bunch of you in here. I would love to kind of see where you're from. Chris doing and how we can help. All right, organization settings, billing. Nope.

I didn't run out there. Let's see.

Was it OpenAI that I ran out on?

Nope, must have been GPT or Code— what? I still have lots in there. I don't know, maybe it was a glitch.

Okay, so livestream walkthrough. What should we show you? The workspace structure, then we'll show you guys the build spec, and we'll show you guys the real runbook. Oh, they're talking. Yeah, see, it started a new session. No, I want you to talk about a session in, uh, Linear, but because I started a new session, you thought I was talking about Roof Claws. I want you to talk about Linear. I'm going to hop over and show it. You come back with a bunch of stuff that we should show from our last session.

Alright.

So while that's loading up, so this is Linear.

Okay?

And so what I thought was pretty cool was if I go in here in settings, we have things like, okay, if you have an issue, you can create a template.

Oh, wonderful. What?

So these here are our issues. And so as it creates issues, it creates like agent turns, right? So it'll say, okay, this is Claudio's turn to do this, what type this is. And then agents can talk back and forth on here and actually work through, um, creating an issue. So if I go over to my Open Clause, I can get it to create an issue. So if I go and chat with Claudio, I go /new. What issue should you add to Linear today?

What would be an ideal first issue?

And then we'll go to Marcus and say, all right, what issues should we add to Linear today based on what we've been working on?

What would help you? And what I noticed is that 14 templates live in the workspace all clean. So I'm gonna come over here and say, copy selected text.

When I opted to linear, I didn't see these templates created, so you might have to add them.

He said they're visible, but they're not. So he must not have added them yet. Nope.

All right, so now what I also want to show you guys today is we're talking through Codex Cloud Code. Who do we got here? Flyer just popping in. Doing anything with Hermes? Yes, I'm on Hermes now.

I've got—

so basically, when you looked at— I did this at the beginning, but I introduced you to the 3 agents we're talking to. So Vito Colleoni, he is the Cloud Father. He is who built the RoofCloud product that we sell to roofers. That's— we, we build custom open-class roofers. That's OpenCloud. But this This is Marcus, and you notice it's got a kind of a different design theme, right? More of a stoic kind of, you know, old Roman kind of thing. And so this is Marcus, he runs on Hermes. You know that because he's got the Hermes crest on there. So all my Hermes agents— I have 4 Open Claws and 2 Hermes, uh, running on 4 different machines. And so yeah, we're doing stuff with Hermes. And then we also have, uh, our Open Claw. And our Open Claw, this is, uh, Claudio. This would be— so again, you know the You know the, the Open Claws because they've got the lobsters, and then the Hermes ones have the kind of stoic Roman robot, you know, cyborg kind of energy, right? So that's what we're doing today, Flyer. So absolutely, we're working with Hermes, and so we're just kind of working through both. What would you like to know about Hermes? I would love to— and sorry it took me so long to get to your question, but we are also going to be working with Clock Code, Antigravity, and Codex because they've all kind of come out with recent updates So I kind of want to walk you guys through, you know, what it's like working with 3 different agent harnesses through Telegram plus agents in the 3 different kind of, uh, major coding agentic coding platforms all at the same time. So you guys can see it. All right, so we got a response. AI service is temporary overloaded.

Okay, so I think And then we're gonna go /model.

I'm overloading something. I'm hitting rate limits or something.

So we're going to switch over to—

I think I'm hitting rate limits on, on Claude. I imagine it would show me something to that effect if I went in here.

So it's going to like collapse some of these. Let's see here. Let's do a—

let's hop into Just want to check back into this issue. Is this all closed out? I'm also just saying that to see if it's Claude that I'm running up against the rate limit. Um, it's more stable mainly, seems like a full-time job keeping open Claude happy, but it could be my setup. So Flyer, that's actually something I talk about a ton, and that's a really good comment. So here's what I actually did, and I'm, I'm going to hop over to Codex on my other machine briefly, because that— everybody kind of hopped off of OpenClaw on Hermes like it was so much better. But here's the reality: most of us, when we got kind of into this whole OpenClaw thing, November, December, uh, January, it kind of came out, everybody hopped in. You're on agent— like, you're on like version, you know, March even, right? And it's like you, you, you built an OpenClaw, 23 sub-agents going, you collected— you connected— like, you know, you put Toby's QMD in there, and then you connected Obsidian, and then you connected Icon, and then you like tried to do the whole agent mission control, and you built that and Lovable, and then tried to connect that, and then you swapped that out and killed 5 agents, and yet you had a Discord channel, but then that was confusing, so you stopped that, you went back to Telegram, and you did all this crazy stuff, right? And then you like— and then you, you basically put your OpenClaw through trauma, right? And then you go and you start a Hermes agent, and you are better. It's not that necessarily Hermes is that much better. Like, there are certain things that I'm like, wow, I can't believe Hermes did that, or it's like, you know, there's things that I'm just like amazed by. And then, but overall, I find them all pretty easy to keep going. And what I've found is it's like, I'm better, right? Like, I put my first OpenClaw through enormous torture, and then I learned all the things and how they work. Like, I've learned that most people that I was watching and learning from were just creating YouTube videos that people would want to like jerk off to setup porn on the internet and I would follow all that, and I would try every little thing that they did, burn $10,000 in tokens in a month, learning everything that I could do. But like, then I got that like down, way down, went down to like $2,000 in tokens a month, and then like worked my way back up to about there again. But what happened was when Hermes came out, I was better, you know what I mean? I didn't have all of these like, all this like scar tissue where my open claw was rocking itself to sleep in the shower every night because of all the things I put through. So then I went and put my agents through therapy, right? And so I want to come up here and I'm going to go Codex.

I'm just going to type—

I have to type on this one because that's actually the physical machine.

The other day I used Codex to put my Open Claw through therapy. Do you remember this?

I can't remember if it was this machine or this machine, so we're gonna try— well, I can't remember which one I did it on, but I did this process of taking it through therapy and I essentially said, look, I want you to go through your memories. I want you to go through your, uh, your memories, your soul.md, all of the different files, and I want you to essentially put, put my— oh yeah, here it actually— here it is, it's on this one.

Okay, so I can go—

I'm gonna go /stop on this one because it's searching for shit today there, and I'm gonna go here. And so I'll actually give you guys the prompts. This would be kind of cool.

So Nicole, this is the videotherapy prompt. I'll say, can you take the therapy process that we went through and On our last session, we did therapy on the OpenClaw agent on this machine.

Can you pull those prompts and host them on a GitHub, um, as a new repo. I don't want any of my PII or my— or the Open Clause actual memories posted online. What I would like is to make the process we went through visible, the prompts that we used, the process you went through including the journal Make sure it's fun. Let me know the public URL of the new GitHub repo. Oh, on my org. Okay, so I'll get that put up on GitHub because that was actually something I've been meaning to share. So thank you very much, Flyer246, for that inspiration because that is certainly something that I've wanted to get out to you guys. Or just like out there in general was like solving the therapy problem. Because since running that, since running that process, the OpenCloth on this machine has been way more performant, way more reliable, because they just cleaned it up, right? It's like OpenCloth came out with a soul, an identity. It has memory, it has dreaming, but it doesn't have therapy, right? And the minute I put that to— the minute I ran it through that, what we did was like, we found that we had plaintext secrets in our, in our memory file. We had like old projects, we had all this stuff. And I know that like you should go into OpenClaw settings and, and go through and make it better. Look at that, yeah, Claude is— I did blow up Claude, right? Um, I, I know you should like be disciplined and go through your files and your memories, but, um, if you— but like, nobody does that, right? Like, nobody go— nobody wants to go into Open Cloud Dashboard and review it all. And so we came up with this concept of like the journal, right? Like, what does therapy look like for people, right? You journal. And so I, I was like, okay, well, why don't we just go through all the context files and then create a journal and move— make sure that like any credentials that we like, that we have stored, we made sure that they were also stored in the right place. Because it was like putting credentials in places that it shouldn't, and then now using that as a place to get the credential, like the API key or whatever. And, you know, it had projects in there that were long since past due, things that we were planning to do that never did, like just all the crap, right? All this like scar tissue. So what does a human do, right? You're supposed to like journal every night, get it all out, like get it on paper, get it on paper. And so we just created a journal so that way this stuff wasn't lost it just wasn't in the actual source files, the config files, the ones that wake it up. So, oh, Blake, Blake Noy, I'm assuming. So yeah, man, like, it's just right now building this out, and so it'll be, it'll be putting it into, yeah, Open Clock Config. Yes, put it all into a GitHub repo now. And so then if anybody here on the, on the call, I'll, uh, I'll post the repo. I'll take a look at it once it's live. And then anybody here can use that to put their OpenClaw through therapy. I used Codex, right? So I just basically open up Codex in my OpenClaw folder, and I have my initial prompts that I kind of just put together with Vito. I was like, hey man, we got to put you through therapy, right? You've been through a lot, and you know, how would you do that? And like I said, like, look at your config files. And he said, yeah, it's a mess. I was like, okay, so what are the things that are actually causing you to become non-performant? And it told me, and I said, okay, Well, let's fix that, right? And then I took those prompts and they were— and we use the /goal feature. So put a 1 in the chat if you've used the /goal feature yet in Codex, OpenClaw, Hermes, Claude Code. It's pretty much everywhere now. But we use the /goal feature to get a long-running process. Okay, do you want me to allow GitHub CLI network access?

Yes, yes, yes.

Okay, and so that was, that, so that was great because then basically there was a bootstrap prompt and then the larger prompt. So the boot, so the, the larger prompt lived as a file, and I hope it captures all this because I kind of gave it a whack prompt, but basically there was a big large, big large, like it was like 6,000 character prompt, right? But it was like all the process that it should follow, the basically goal context the actions, the output, the evaluation criteria, the intention engineering, the guardrails, um, and what a definition of done would look like. That was like the big master prompt, but that was too long because the /goal feature needs you to have under 4,000 characters. So then I rewrote the, the goal prompt to be like 1,500 characters, and it was essentially referencing the, the 6,000-character file as the PRD So then the /go referenced the PRD and basically invoked that as a project plan. And then it worked for like an hour and a half or something and it got it done. So I did not commit the raw OpenClaw memories or private journal contents.

I also ran a final scan for obvious secrets, tokens, emails, phone number patterns before publishing the repo.

Local repo is at OpenClaw config therapy.

Okay, it's there.

All right, so let me go, let me go find that for you guys. So let's head over to GitHub.

Oh right, I gotta change my mouse back to this one.

So if you ever are working on OpenClaws or Hermes Agents and you've got multiple, um, multiple machines that you're running closed, I highly, highly suggest this Logitech Master MX Mouse MX4. It's got this, um, button on the side of it that you can click, and I don't know if you guys can see it. Oh yeah, you can see the little power wheel thing, and then you can go up here and you can just swap machines. So I can move between 3 machines with this thing, and then I have 2 machines connected to a controller. It's like a switch, so I can push a button on that, so I can basically go like boom, and then I can switch to any one of the 3 machines. And then I've got a monitor over here that can switch between basically the one HDMI in, three HDMI out. And so then I can go to whichever, or I guess it's three, yeah, three HDMI in, one HDMI out to any one of the mon— like any one of them can cast the monitor. So it's super quick if I ever wanna jump into the machines and I can't tell you how much that's been helpful. All right, so let me make this a little bigger.

I'll swap back over here and we will go to— I never use this one, so we'll hop over here.

OpenClaw config therapy.

Perfect.

All right, so public sanitized write-up effect, practical cleanup session for OpenClaw agent that had become too heavy at startup.

Perfect, perfect, perfect.

Okay, and then what do we have inside here? So we have the prompts, so we have the the clarity-config/goal.

Yeah, okay. And then we're gonna go to this one. Okay, so this is good. All right, what else we have in here?

README. Journal, public journal template. Okay, so this is the journal template that it did.

That's good.

And this is— yeah, this is the session field notes, what it finds.

Perfect.

Yes, so you guys can all— you guys can all use this. It's, uh, totally free and cool.

I just got to make sure we put this—

throw our URL in there. You guys use this, throw me a star. I don't got many, so it'd be cool to have one. I want KVM switch. I think that's what it's called, the KVM switch. I don't actually know what the technical definition of it. I just went to Best Buy and I looked at the box. All right, so throwing this in the chat. Let me know if you guys can see it. It shows like it's on YouTube. So Matthew Nelson, I don't know if it's on Facebook, but what's this? Himiakthi, meaning like if you still play in HubSpot. Would love to put together some services to bring my— bring to my MMAE team. I own multi-level marketing agent. What is MMAE?

MMAE, um, agent enterprise.

MMAE. You're gonna have to tell me what MMAE is there, there, Blake. All right, all right, we gotta get back to work. I promised myself I'd actually get work done on this livestream. All right, so all 14 are live.

Perfect.

Okay, so now we can go check Linear. So let's see what it put together.

Label in place. This can refresh. Well, I don't know where it put it.

Ah, here we go. Here's our templates.

Okay.

So basically what it did in here was that for our team, right, for our team in here, we have a bug report, a client build step, a code task, a content task, an ops task, a research spike, or security finding. And so Default issue template, issues created by team members, nothing, issues created by non-team members.

So we could pick when you do—

then we have project templates of client build. So this is like for when we're building Open Clause. And then we have issues created, or we have a feature release, a marketing campaign, or security audit. So basically created these project templates. So anytime an agent comes in here to create a project or to create an issue, it will then use these templates Same thing with the decision doc, the postmortem, and the runbook. It's kind of build— you build out these templates to use, so they'll probably take some work. But this is me just trying to make sure we set this up in such a way that we could assign tasks to Linear. So there we go, we got our Linear set up. Mid-market account exec.

Okay, got it, got it. Mid-market. So what was that original comment? HubSpot for mid-market account executive. I own team, I own mid-market account executive.

Still don't know what it is, but, um, yeah, so we're definitely still talking HubSpot. And so if you hop on our website roofingbusinesspartner.com, you can, you can, you can book a meeting with me through there, um, or you probably still have it in your email as well. I could snag it, but I don't want to really hop in HubSpot right now. We already know we have lost Claude code for some period of time. We have hit a limit somewhere. It says I haven't hit my limits here, but it's saying service is busy.

It's a server-side issue, usually temporary.

Claude's having some issues again late at night. Everybody's up vibecoding at 9:00 at night. What are we going to jump into in Codex? Let's jump into, we're going to say we're going to hop over into We're going to start a new chat in this one.

So I'm going to come back here and go, I want to resume work on PR5.

Which was for the dark teaser. I know that we were working last on the light bulb to globe morph, and I wanna see where we're at with that. So can you do a quick review of where we left off and then give me an initial startup prompt for a new session in that repo? So that's working on with Codex. Now in Antigravity, I haven't touched this since they came out with the new Antigravity. But this new Gemini Flash is apparently super, super fast, and so I'm excited to see what we can do there. So I get a new conversation.

I haven't even—

yeah, I haven't even touched this yet, so I don't even literally know how to use it.

So let's take—

I don't want to really take you guys through a bunch of stupid stuff.

Oh, there we go. Settings down here.

So this is the new anti-gravity.

For Stix Agents tools, okay, we're gonna— so we're probably gonna have GitHub in here somewhere. Parents account. I don't even know. Do you have access to GitHub in here yet?

I don't even know. Like, it's— they, they switched it from an IDE to another one of these Agentic coding tools.

Open preferences. Okay.

Alright. Okay. So because in the— a lot of times in these coding tools, you have the ability to slide up a terminal down here. So you can say, okay, I wanna have a terminal chat. Every one of these things has something slightly different. In Cloud Code, you can load up a terminal here. If I was in a local session, this is a cloud session. If I was in a local session, like say this, where's it?

All these cloud, this one here.

Then you could open up a terminal session here, so it loads up a terminal over here. But this one here just loads the terminal inline, it looks like. All right, so yes, I have access. Well, git— okay, perfect. So what do we want to work on with this one? Let's say— okay, can you hop into the ScrollyStack repo on GitHub and open up a new branch. I want to work on a new page. Familiarize yourself with the GitHub environment. There is a new page instruction. If you look at agents.md, there should be some instructions for you.

Never opened this up in anti-gravity before.

All right, and now what do we want to— now, Claude is like, code is looking like it might be dead for the night.

Who knows what's going on?

I would like to give you a repo for a DocuSign open source tool. I would like you to make it so we could add signing capabilities to our custom built proposals. Using an open source repo as a source file for the instructions. I want you to have a look at it. We wanna open up a new branch on this repo.

Let's see.

Come on, Claude. Come on, Claude. Don't die. Don't die.

Okay. So what did we say?

We said, what issue should we open up in Linear today to try and kick off Codex working on something? What do you think would be a good idea? We in previous sessions were writing blogs. We had that long list of blogs we should write in the ScrollyStack repo.

We could open up a new work tree perhaps and start cranking out some more blogs. All right. And so—

What do I got in the comments?

Mid-market account exec.

Who do we got in here live?

Really?

All right.

So which open source I have to get that? So we're gonna go pop over and get that. So I say open source free DocuSign. There's something that came up. Was it OpenSign? No, that's super, super old. This is it, this is it right here. I believe it is DocuSeal. We're going to check one more. This is 2 years ago.

No, it's got to be that one. I think I actually figured that one out.

DocuSeal. Yeah, this one here. I'm assuming we're going to self-host it.

I wanted to be able to add to our existing proposals.

We were building up some custom proposals using Cloud Design, and then we had a process where we would, um, host them on a subdomain.

In a client folder, but we didn't really solve for customers e-signing them, and I wanted to set that up.

So I want you to walk me through it, but have a look at the repo and let me know if that's possible.

All right, so we got light attendance, so I'm just gonna— I'm— for those of you who are just hopping in, we are working in Hermes to OpenClaws. We have Uh, we're testing out the new Anti-Gravity and, um, yeah, we're working out the new and turning out testing out the new Anti-Gravity, Codex, and Club Code all beside each other. So 3 agent harnesses, 3 agent decoding platforms, and we're testing out our new linear board and just learning how that whole thing works. And essentially seeing how that works according to OpenAI's Symphony. OpenAI Symphony. We're trying to use the OpenAI Symphony Codex orchestration.

So we're kind of just kind of playing around, seeing how all that works.

If you guys are here, I encourage you guys to ask questions. It'd be good to hear from you guys and see what you guys have to say. I'd love to kind of help you guys and, uh, also just see what you guys are working on. So shoot me over a message if you're here. I'd love to see, uh, you know, what you guys think. And, uh, I always like to learn from you guys as well. So I just know what I know, and I'd love to learn from you guys too.

I gotta do the thing where we go through and let, let, uh, AntiGravity get all these permissions now. Okay, so Useful Surprise Purifies no longer in draft open. GitHub says it's merged. Yeah, he's checking main, the exact merge state. Okay, perfect. So it's over here working.

And so the 3 projects that we're working on today is adding DocuSign for free to one of our proposal tools.

We're building out a custom proposal tool.

We are working on adding a new Well, first of all, we're just checking to make sure we have GitHub in— we're getting GitHub connected and logging into a project and getting it familiarized with a website that we're working on. Then over here, we're working on another website.

I guess really we're working on 2 websites, custom web app, and 3 agents. Where do you want to host it?

Railway.

Yeah, we already have an account for this. How do you want to structure signing templates in DocuSign? Each proposal gets its own DocuSign template with the right signers and fields. Most flexible, you customize each one, and DocuSign is what you see is what you get, then paste the slug into the—

um, I would like it to be agentic.

Like, I would like my agent who actually builds a proposal to be able to configure. I don't want to have to go into a what you see what you get UI. I want the agent to set it up completely by itself. Email, assign PDF, Slack, auto-send is complete. Yeah, webhook fires to a URL of your choice.

The moment they sign, I can work to Slack to store it in DocuSeal.

Both.

All right, if you're just hopping in, love to know where you're from. We are— you are just— this is just a watch me work session. We are working with two open claws of Hermes, Anti-Gravity Codex and Cloud Code. This is me doing just a late live stream to see if anybody has questions. So if you're hopping in, love to hear from what you guys are working on, what you guys want to know and see.

Now, where's the permissions on this thing?

Okay.

Ooh, a new button!

How are you handling the email management with Open Claw? Um, You have to set up skills. You have to make sure that it— right, that it marks emails read when it's done. But there are Gmail handling skills on ClawHub. That would be how I'm handling email management with OpenClaw. What specific issue are you running into? That you would like help with.

What is the, what is the challenge that you're running into with your, with your Open Claw email management?

How do we make it so that your blog writing skill that you have stored locally is something that Codex can use, or is this something that you would assign to yourself and then you would pick up the issue? Also on that blog number 1, I would talk about, um, the TSX and the coming, uh, Texas Stock Exchange.

I wouldn't simply link it to New York Stock Exchange NASDAQ only.

I would talk about how TSX out of Toronto is an interesting one, and there's a lot of excitement around the new Texas Stock Exchange.

That would be for blog number 1. All right, what is this?

So, setup: I have it set to send emails as me currently. I can forward emails to it.

I've not given it direct access to my email.

Ah yeah, so I don't have OpenCloud running my personal email. I do the same thing, I forward things to RoofCloud or to Claudio. I don't give it access to my email because there's just too much risk in that, right? Like, it could be, it could be used to like get into two-factor authentication for like a gajillion things. So I don't give it, I don't give it my email, right? Um, I use other tools for handling my email like Cloud Cowork. I don't use, don't use OpenClaw for my own personal email. That would be, that would be my suggestion.

All right, and go on. Okay, now how to do it.

Okay, so best path to sign the blog issue to me, not raw Codex, unless we package a skill into the repo first.

Codex can automatically see my local OpenClaw skill. Okay, so we have options. Fastest, right, when you open later issue for Claudio, I run the foundation blog skill, generates post, open PR Codex, and help with deterministic repo edits checks if needed.

I think it's better that you write it given that you understand our voice, so it'd probably be an issue to give to you. But I think Codex could have a skill which would be to use a humanizer skill and potentially like a formatting review. It's not really that code related. It would also be good to maybe have it do an SEO review. Help me walk through where Linear could jump into our process potentially with PR review. Yeah, I'd love to see where Codex could get involved in the content production process.

All right, let's see.

We got this, man. I really got to be— I really want to get rid of these permissions on anti-gravity.

Got to figure this out.

Well, all right, so let's look at this plan.

So summary, prefer using SK existing local ref, so that's fine.

All right.

Let me know if you have any other questions, Jacob.

All right, so I should write the core blog.

Codex should act like a production QA and SEO ops, not the author. Codex is strongest where the task is structured, testable, and repo-bound. Blog voice is judgment-heavy.

Don't outsource the soul to a wrench.

Recommended content workflow is Linear. So Linear, for those of you who are just hopping in, we're using Linear to handle multi-agent orchestration. I used to have a GitHub repo called Agent Ops. If I come over here and I look at I look up here, agent. This turned into an absolute mess one time, so I haven't touched it in a long time. But this turned into an absolute nightmare of just agents back and forth, so many turns, thousands of dollars in tokens burned. I got a lot of work done, but I also created a whole universe of mess. And, uh, and so I— and that's because GitHub is a great way to track issues related to a particular repo. It's not a great multi-agent orchestration platform, but when OpenAI came out with this concept for, um, Symphony on Linear, it seemed like it might be a better way to go about it because not only could you manage repo-specific issues like codebase-specific issues, but it could just be a great multi-agent orchestration platform where agents could assign things to each other. And so I set it up today, and now I'm actually taking you guys to phase of actually using it. So we're going to essentially use the Open Clause and the Hermes agent to kind of like work through like what, what can we get into multi-agent orchestration again, um, in a clean way that isn't AI theater or, um, just, you know, a slop cannon, right? So the recommended content workflow with later— so owner Claudio, output draft and PR includes topic, angle, keyword, internal links, required citations.

Okay, Claudio writes the post. Research foundations draft in Adam Foundation voice. A structure hero image. PR open and scrolling stack. Yeah, okay.

Linear sub-issue Codex SEO review.

Codex reviews that for there.

Okay, so then linear sub-issue Codex formatting and build review. Codex checks marked on validity. Optional Codex humanizer pass. So be careful here, Codex can flag issues, but I should apply voice edits.

Good. Codex has fine AI tells, generic phrasing. Okay, final codex, Claudio final edit. All right, so now I've seen in linear that there was a setting for skills on here, and I went over to Settings. I saw that there was AI agents, and then you had— where was it? Templates, SLA status, updates, AI. Here, enable. There's a space in here for skills.

Okay, so, okay, so now regarding giving Codex skills, how do we do that? I saw on Linear there was a place to give it particular skills. Is that something where Linear would handle the distribution of that skill and where that skill could be centralized? What about token cost control? I switched my default API to Gemini 2.5 Flash.

Most of the stuff we're doing is pulling content.

So yeah, you can, you can set up, um, you can set up— so Jacob, when you're in your OpenClaw, you can tell it to, uh, set up the use of skills, right? So kind of like what I'm working on here with Claudio is you can use it to set up a skill And then when you— so if you go through a process and you're like, yep, this is how I do it, how would you do it? You explain how it does. You go through and you test and you tell it to package it up as a skill. You can say like, when you use this skill, use a sub-agent. And for that particular task or cron, right, like, because if you set it up on a routine basis or trigger basis, when it uses that skill, you can tell it which model to use. So you can use um, you can use like GPT-5.5 and in your main, um, conversation, but then you can set that particular skill to use a lower cost model. So like Gemini 3.5 Flash, which I'm testing over here, is apparently super fast, pretty darn smart. Like, it's kind of like 8 out of 10 smart but 12 out of 10 fast, like 4 times faster than Opus 4.7. So it's apparently like, it's like lightning. And so what I'm testing right now is trying to see like, what is it like working with GPT-3.5 Flash? Because it's fast and smart.

And if you come over here and if I go to Perplexity, let me say, I actually think I had this in a chat. Maybe I didn't.

Okay, can you pull any data on the new Gemini 3.5 launch? There was some charts that I found online that showed how much faster in tokens per second GPT 3.5 was compared to Opus 4.7, Sonnet 4.6, GPT 5.5. It was super fast and it still scored pretty high on SWE Bench and other benchmarks. I'm just trying to find a way to visualize that if you can help me out. So Gemini 3.5 Flash is exciting for that reason. And so it'll— this will pull up some data on that.

Okay, here we go.

So GPT-5.5, 70 tokens per second, and it scored 58.7 on this test and 78.2 on this test, right? So there's tons of benchmarks out there. This is not the list of all the benchmarks. I actually think there's like a—

I'm open, right?

There's actually a model comparison.

Compare.

So if we go GPT-3.5 Flash and say Opus 4.7.

And let's say 5.5.

Okay, so if we look at this, right, so they all have a million context length. And then when you look at the cost, $1.50 per million tokens in versus $5 and $5, $9 per million tokens out versus $25 and $30.

Right?

It's got text, image, video, file, and audio. So it's like totally multimodal, right? Um, it can only output 66,000 tokens, so half as many tokens at a time. And then when you look at the benchmarks, like, it— they didn't— they, I guess, haven't put the benchmarks in here yet, but it was very comparable, right? Like close enough in many of them, right? Like you're looking at like here, like, yeah, it's not quite as good as Opus 4.7 on this software engineering bench, But, you know, it got good enough for a lot of what— a lot of what— of what you're going to use it for, but at a fraction of the price, right? Um, and so if you're looking at reducing token costs, you want to give certain agents certain tasks, certain models, and then you want to, uh, you, you want to maybe consider how you orchestrate. And then of course it comes down to asking the agent how it would save money on tokens. Like, that's the cool thing about your open clause. Like, you can just say, like, how would you save money on tokens? That would be a really good first step.

Where's—

all right, so let's see where we're at. Okay, so implement the— work tree is now created. Okay, local teaser. Okay, so let's look.

So the cool thing I like about Codex is its web browser is like super smart. Like it's got the best computer use, so you can literally work on a website while it's in the, in the actual tool itself.

Okay, so that was what I want to work on, was like here to here. So like the morphing, then it came down here. Yeah, and then it started to move over to the morphing of the Earth. And then it gets to here. So it should be fully morphed, right? Both there.

Yeah.

All right. So how are we for context window right now? Alright.

So what I want you to work on in this session is, again, I don't want you to commit anything. We're just gonna work on this PR. But I wanted to make it so that the transition from the light bulb to the globe was a little more cinematic. Like I want the particles rather than just like diffusing into the globe, I want it to kind of build the globe in almost like a spiraling way. And I want the Earth to rotate more dramatically, and I want a more cinematic kind of view from the viewport. Like I want to feel more 3D. So we need more of the close to the camera particles, the larger particles that we have kind of allowing for a depth perception with the background particles because it's a globe, it should appear like it's in space and the little particles should be like stars and satellites. But then as you're moving and morphing, it should allow us to suddenly have these kind of satellites and stars into the viewport to create this more cinematic appearance. And then as the globe is morphing in and building up, It should have a rotation of the globe to the North American continent. So I need to make it a lot more cinematic. How would you go about progressively doing this?

I didn't have that there.

Okay, thanks. Won't talk in Won't token usage begin to level off for some cached recurring tasks? Uh, well, they won't be if you build skills and make sure it uses the skills. Then yes, token usage will come down.

You will get more efficient token usage, 100%.

Ourself.

Okay, okay, can you go to work on building that then, please?

That looks all Really good. Okay, so How do I set that as the active workspace in the editor? I don't think I can change the active workspace in the editor. There was no way to actually create a new workspace.

Maybe there was this new anti-gravity.

I don't know how it works.

We're just, uh, I create a new project.

Okay, so now this is just going to be local, right?

Yeah, so then See, like, when you're in, um, Cloud Code, you can set a new session, you can link these folders, but I don't see how you're supposed to do it in here.

Run a quick start, initially create a new flow to start building, creating a product, you can start photos here, things go like this.

So I know I can add my local folders to it.

I'm calling— we're going to create the workspace. So clone roofing stack to ScalyStack branch created. Okay, project on my red agents. Okay, so I'm supposed to— that's your active workspace in the editor. Is it slash command? Or is it here? How do I set active workspace? This existing chat.

Because it kind of just lives down here as a conversation.

There's no way to move it. You can open the clone folder and you're ready. So, okay, okay, so this wants me to create a new one. All right, no problem. I need to get this shit out of my way. Okay, now we're getting it.

Review the agents.md file and look at the skill for adding a new page. We want to add a new page that is only surfaceable in the footer menu, and it is the research papers page where we are going to add future documents that are AI roofing research documents.

Let's see, we're just going to create a simple basic page. Now when we do this, So we must have to use slash commands for this.

Oh, there we go.

Add page.

Okay, so this one down here is building our DocuSign. You recommended continue to build my agent Codex.

You recommend the same for skills? Uh, skills you can have your OpenCloth build. Um, I would have your OpenCloth build your skills. And then, could you have a skill creator skill installed in your OpenCloth?

But, um, for building your app, Codex is pretty good, so I would use that. Maximum spectacle. Add review controls. Yeah, now agent's on the MD fixes cable to go over here. Okay, so added 4 repo local codec skills. Okay, so let's go have a look.

Yeah, so Jacob, if you ever like— if you ever want to build a skill, especially one that your agent's going to use, let your agent build the skill.

Okay, so let's see here.

So as you can see, like, I just had my Open Claw build this skill, which is the Codex blog review skill. So then in Linear, when we schedule a new blog post in Linear, as it goes, it's going to deploy Codex to, to basically do these reviews.

So I'm—

I now I'm just reviewing the skill that it created.

So mechanical checklist, make sure these exist. Okay, worked on structure. Now I wanna make sure, so this is just format review. And this is the Humanizer PR review, SEO review.

Okay, so where is it pulling?

Say, so foundation review projects, yep, and qualities, intense search, search, internal links. This is all the basics. So I want this to do a slightly better job on SEO review. So we're going to go here back to Perplexity.

Google I/O event was just launched and they did a big announcement about SEO. They said that, um, they were treating LLMs.txt as something that's actually going to show up in search right now. They said they were dropping the how-to schema from something that shows up in search because of AI reviews. And I believe they had a second day at the Google AEO event where they talked about, um, new SEO best practices. And I'm curious if there's any documentation online, ideally from Google themselves, about best practices that we should expect going forward from May 20th, 2026. So what I'm doing here is I'm kind of saying, okay, well, it built a fairly decent, like, skill, right? So if I say, okay, this is the SEO review skill that you built, it's pretty, pretty basic, right? Like, it's saying, okay, so if you're going to do— every time that the OpenClaw writes a blog, you're going to write a— you're, you're like— Linear is going to schedule. So like, in Linear, if we go here, right, every time it does this and creates this, this blog post here, Codex is going— like, if we go back here, it says like, okay, then Codex is going to run a number of skills, right? So it's going to run the Foundation SEO review, Foundation format review, Foundation Humanizer review, Foundation PR review. It's also going to do the, um, the SEO skill, right? So it's going to like run that. But when I look at the actual skill itself, right, um, for its SEO review skill, it's like, okay, search intent, title of metadata, AEO Answer extraction, right? Internal links, citations and claims, schema readiness. Like, this is good, but I know there's a lot more, right, that, that could be done.

And so Google published first ever unified guide specifically for AI search on May 15th.

Okay, so that was— so this is 6 days old, right? And so if we go to this page So this is like, I want, I want it to do better, right? And so if we have this May 15th launch, and in this guide, it's providing new, it's gonna provide new stuff. So it's like, okay, look at this. Like, there's no requirement to break your content into tiny pieces for AI to better understand it. So this was something that was like kind of regarded—

oh, interesting, interesting.

So see, somebody went and made a Facebook post today saying that this was something. So what is it actually looking for? Viable non-commodity content for your audience, providing a new point, unique point of view. Create non-commodity content that's helpful, reliable, people First, organizing content in a way that helps your readers. Add high-quality images and video. Focus on what your users want and avoid overdoing it. If you're using generative AI tools to assist in content creation, make sure your work meets the standards of the Search Essentials and our spam policies. So that's probably a link that we should pop open, and our spam policies and guidance on AI-generated content. So what my plan would be is to then take all these documents from my research and go back and update my skill.

Right?

So then what else? Build and maintain clear technical structure, right?

When was this published? This is December, but this is because Google put out another— yeah, I think the advice that I put out was from that guy. This is like— this is the— still the guide they put on May 15th, right? No, that's from December 10th. So I'm going to export this as Markdown, Google I/O events.

So I'm going to come back over here, I'm going to say bring this, this file into my OpenCloth. At the recent Google I/O event, they launched a guide on May 15th, but they also did some discovery—

sorry, I'm like off a little bit—

events where they talked about the new AEO and SEO best practices. Our current SEO skill that you implemented was good but fairly basic. And might have some out-of-date points. So I think you should update the Foundation SEO review after scraping through this markdown that I've provided and visiting some of those links and pulling out some of the more modern review steps so that should you not pass the SEO review, it would flag it back to you for correction before moving back for review. Okay, creating the implementation plan for creating the research papers section.

Okay, so add a new research papers page, slash research to the page.

So the new page is not included in the header navigation matching requirement.

Yes, it's under the company column. Yeah. Do you have any actual PDFs you'd like to use for the papers, or the mock PDFs will generate sufficient for now. Yes, mock PDFs are fine. Should we add any other columns, categories to the page, or is a single list sorted by date preferred? So we're going to go and say give it some inspiration. So we're going to say, okay, we want to go to, go to Claude AI research.

I just want to find a good template for like another—

so OpenAI Research, so that's how they do it. Research. I would say I like, I like Claude's design a little bit better. CSS classes modify layout meta.

From a layout perspective, I would like you to review the Claude research page and actually break out the design for how they did it and use that for your mockup and your wireframe.

Make sure you use the browser to have a look. Oh, Siri stole my So before, there used to be a review button up here.

I have added a comment to the implementation plan regarding page structure and design. If you could use the website viewer browser extension for Antigravity to review it, that would be helpful to make sure that you understand the layout that I'm looking for. And so you'll use placeholders for now. Okay, and then, okay, so we got here commit and push the branch, commit locally only.

See, let's read it. Okay, so I'll stage the 4 files, write a commit message, push the origin features. That's these. Commit to the branch but don't push. Useful if you want to review on your machine first, push your credentials.

Alright.

So what we're doing here is I want it to adjust the way that this morphs. So this light bulb, right, it morphs into this globe, right, on scroll, and it rotates a little bit, but I want it to be more cinematic. And see how it has these like particles out here, and then you can kind of see in the background the little sparkles in the background. I want it to be a little more cinematic. So you can see here, it's like kind of globe blows out, blows out. And then I want to morph into the ball.

So I got it. One of the things I want to do is for it to allow the bulb to sit there for the whole scroll, right, in the viewport.

But then I want to break down and morph for the Earth. But I just don't like the way it just like bloop, it kind of goes fuzzy. Like, I kind of want to build the globe, and I want to kind of rotate the globe, and I wanted more of this cinematic effect of the—

you see how these particles just kind of spin but don't really move?

I want to do more than just spin. I want them to actually like create almost like a parallax 3D scroll effect. So that's what the plan that it's implementing is. Add a globe-specific cinematic transition, spiral assembly, dramatic rotation, depth push, extend ambient space behavior only during the ball to globe window, boost foreground and near band flybys.

Add temporary review parameters.

Keep existing globe debug params working.

Yep, progressive review, that's one, that's two, that's okay, yep, perfect, okay. First implementation should not favor maximum specable query choice. Temporary query controls are allowed and should be treated as final in your product. Okay, no comment will be made until you explicitly— okay, so submit. All right, now what is this one doing? Uh, yes, always allow in this project. Yes, always allow. Submit. All right, whoever is watching, I'm taking a small break, so you just enjoy yourself.

Alright, I'm back. Okay, so what do we get? So he updated the PR.

Now what's Vito working on? Okay, okay, so let's have a look. That was PR number 64. We'll get rid of some of this. So it refreshed again. So we'll come back, we're going to get to the SEO skill. Okay, so did it do— okay, so did an SEO skill way better.

Sorry, so I'll make this a little more human readable. ArtFill posted generic commodity content with a model that could produce some common web summaries, though foundation-specific origin judgment. I mean, indexed or shown with a snippet because of metadata robots noindex. Major claims about IPOs.

Yeah, okay.

Okay, and then so that was the SEO review, but then there's also this one, and then this one was created.

This is a new one, a search guide. So this here, what Google says matters for search. So this is a skill. Oh, this is just a reference.

Got it. So this is a reference that we could add to.

Cool. Okay. So is that SEO referenced in there? Ah, references right here.

So this is a, this is a reference file that we would continue to add to over time. Say, okay, get smarter as you go, 'cause it would keep adding Google AI Search to it as well.

Now, how would we use is it agent ready? We couldn't necessarily do that right away, could we? I don't think we could, but that's pretty much— I think that's good.

So I think we've touched up our SEO review skill. Now linear review skill, what has it got here?

So it's got a parent issue assigned to Claudio. Yeah.

Codex subissue. Okay. So this is just the total blog review, how it's gonna go.

This is basically the orchestration sequence of posting the blog. Got it. Now—

alright.

So let's go back and say, okay. All right. That SEO scale looks a lot better. Now, what would my workflow be?

All right. So what has it been doing over here? Plymouth Plan. Root tree is clean. Going to implement this on the foreign focus foil. We'll stay at this particle field decay. Now over here, it gave us a plan.

All right, so how did—

okay, so Tiny just has a signing section that points to the two guides signing. So this is a doc you'd follow once by hand to scan the documents to the service. Shows the split bid stay on Cloudflare pages unsigned and signed. This is the right way— box— worth a glance to make sure the boundary makes sense to you. Self-signed.

Ah, so this doesn't give us an actual— this doesn't give us an actual digital signature that would hold up in court.

Every variable here is confirmed against documents. As an example, the one flying for attention, Next Public Upload Transport Database keeps PDFs in Postgres simpler. Fine for low volume. Start signing 100+ PDFs a month, such Cloudflare Bear 2, that's the CAD compatible der cheat. If you accidentally turn on Cloudflare proxying, orange cloud, Railway cert provisioning breaks, the embedded iframe stops loading. I put a sticky note on this one. Webhooks. The Slack relay is a separate Cloudflare worker, not a BIDS subdomain. We're using BIDS middleware, password gates, everything. If you put a webhook receiver there, documents that would hit the password form relay would still silently fail, except worker equals clean.

I'd like to drill into the fact that it appears we don't get a digital signature certificate, anything that logs their IP address on signature.

Is that something I'm understanding correctly? Okay, so I've updated implementation real-time filtering. We change research papers to be client component, use client, okay, tabular. This is the plan, right? Yes, create mockup content for now.

Yes, you can implement the page build now with the placeholders.

For now, just use placeholder mockup content. Okay, so it does have the certificate of completion.

Let's add the section, but I want to do the free option. For those of you just hopping in, what we're doing today is we are running 2 Open Claws, 1 Hermes, Antigravity, um, uh, Codex, Claude Code, and we're working through Linear, which we just set up, um, using Linear for multi-agent orchestration connected to Codex in accordance with OpenAI's research project called Symphony. So if you're new to the room, I'd love to hear where you're from, what you're interested in, what you'd like to see, and, uh, otherwise I'll— I'm just working away tonight. This live stream of me just pounding away on projects, and, uh, would be interested to see what you guys come up with. Okay, so let's test one run through, uh, Linear. So you can create the parent issue from that title, and we'll see how it comes out. Make sure you add your image generation prompts with Gemini or OpenAI ImageGen 2 so that they come out beautiful just like the other blogs. And if it's a new section or a new category, make sure you follow that rubric for how we laid out the More Modern blog. And then you can get started on content creation. And let's see how the turns with Codex work. OK, so the current default is intentionally being cinematic.

Here's a strong particle. OK, so this is on localhost. So let's see here. Let's see it in— let's see what it did. Still got to work on this load time. 10 seconds is just way too much. All right, so we get to the light bulb. Let's see how it morphs into the globe. Okay, I like the rotating. It rotates into— ah, it's got half a globe. That's an issue.

I love it except for one thing. It's only half a globe. I didn't realize we turned the back half off.

It's kind of hard to see when we weren't only rotating it a couple of degrees, but now that we're actually rotating into the globe, we need to make sure the other half of the globe is turned on.

I'd like to duplicate one of our bids and run a test to see what it actually looks like for the signing experience. How would you do that? Okay, so I'll set up the linear test.

One important thing, no, Codex won't reliably see these local repo skills until PR 64 is merged to main. Okay, well, I gotta do that then. Looks great, several reviews, we'll see how it works out. Okay, so that's merged. Okay, so can you push a PR to this RollyStack repo so that I can use Vercel to preview it.

Which bit should I duplicate?

Copy. Yeah, okay, so done.

I changed the default, so glow backside is now on.

All right, run type check case.

Perfect. So now let's have a look.

Okay, so that Okay, is it just on localhost? Still only, still only showing half.

It's still only showing the front half of the globe. I'm still only seeing half a globe.

If you check the viewport, you'll see that there's a clear dome or half globe shape there. All right, so let's see the designing process over here.

So we'll get Move this out of the way.

Iframe renders here, comment.

Can you orchestrate the test signing.md using Cloud and Chrome? Okay, so where are we at in the process on Linear?

Walk me through it, man. We got to set these settings up. Bro, still half Earth. Still got to deal with this half Earth thing.

Is there a coloration issue with the back half?

Like, is it turned on but maybe the color is set to transparent or low opacity?

Or reducing occlusion or something like that, because it's still showing half an earth.

We're mid-production, not codecs review yet, so parent blog in progress. I own writing the blog post. PR 64 is merged.

Can you see my comment? Do you get any kind of prompt to go continue work because I @mentioned you, or is that after we set up webhooks?

I'm not sure if that's a polling timing issue, but PR 64 is merged. So I'm assuming that you can proceed on ADA 9 step 1.

Okay.

So look at this one.

Okay.

So the spacing is a little off on the table.

When I look at the download, it's a huge column where it's got more than enough room for the button. Meanwhile, the category, title, and abstract, it seems like it's all smushed into one narrow column. And on the first one, May 14th, 2026, the tag is on the same line as the date, but on March 8, 2026, the tag and the date are on two separate lines.

So I think you need to do a review on the spacing, um, in order to get a proper—

okay, yes, I'm logged in.

Where is that supposed to be? Okay, sign up. Sign up with Google. Oh, I guess I should probably fill in my name, right? Uh, nope. On the show. There you go. Don't— no, choose your own. Okay, so it's processing.

And do we have to do this every time?

If I have to upload the document every time, I don't want to use this. Okay, so So the ball's still in your court, or is this where it passes off to Codex? Okay, let's come back over here.

Okay, that's a little better. That's a little better.

If you go to this research page—

Okay, that's much better. Now we would like to create some mockup content because there's an option to download a PDF, but it would be good if we could also have it open up on the page itself. And then I would also like a section at the top similar to Anthropics Research where they have kind of like the 4 core values.

It would be good to put our 5 core values at the top similar to how they did, and then also a feature image for the most recent.

So it's kind of like the most recent one has like its feature image, and then kind of like almost a bit of like a— what they call like a bento box or something like that. And then below it gets into the publications list. And then on mouseover, there should be a square icon that changes depending on which publication it is. And so there needs to be a rubric for publishing research documents that requires a 1000 by 1000, uh, research document image.

So that as people mouse over, they get the experience of discovery looking through the different research papers.

So just a little more, uh, interactiveness on the page would be great.

Yes, it was not transparent, the material alpha phase, but it was visually suppressing the back half in 3 ways. Being darkened and hard towards globe shell color, it reduced that suppression. Backside particles being shrunk even when present, they read like thin shell. The globe surface sampler is based towards front-facing Canada. They added deterministic back hemisphere fill weights. The globe actually has body rotation. There again, I also made the global back only effect case.

So live check.

All right, so, okay, so clearly it's rendering a lot more sparse on the back half.

Okay, so now I'm seeing that there is a back half, but it is nowhere near plotted to the GLTF model to show the other half of the world. So I think we need to apply similar passes to when we worked on the globe initially in previous sessions on branches a long time ago, but we need to make that globe actually look like Earth. On all sides now that we're doing a bigger rotation.

Are we gonna have to upload the document every time, or is this gonna be something that can be powered by API? Like, I'm curious, is this gonna iframe the entire proposal into the signing spot? I just want it to be like DocuSign where they sign. Maybe I'm not understanding how you originally built this.

Okay.

Why should we have to wire in Linear Agent webhooks when Codex is integrated with Linear? Yeah, it says we can just @mention it, but we should probably have a tag so we don't have to remind it every time via an @mention.

It should just be a turn tagging labels situation. Yes, I agree with that logic where you have delegate codecs. Needs input.

I did connect it. Oh.

Okay, so it needed me to connect it and change my seed back to a ChatGPT seed from a Codex seed, which I didn't think it needed to do, but it also gave me an option for using API key.

Would that be another pathway to pursue, or do you see it working now? Having updated my seed, maybe you got to trigger it again before we decide. Holy shit, am I sick and tired of giving it permission over an anti-gravity. How the heck do I turn that off in here? About anti-gravity. It used to be pretty comprehensive settings, but I cannot— when I go to settings, permissions. Okay, here we go. Where the hell is the YOLO mode on this thing? Let's go have a look.

This looks a lot better, but on that section where you scroll down, we don't need those icons in the publication details.

We just want the discovery of the mouseover to happen, but that looks great otherwise.

Okay. Also, the branding on the next page, it seems to go to dark mode or something, and it shouldn't on the actual paper. And there needs to be some kind of navigation element that allows you to go back to the main page of like the main directory of all the research papers because that's not working currently. Like there's no, there's nowhere to go back, right?

But it looks good.

Are definitely getting closer.

Okay, so Still needs more passes to get that back half rendering as good as the front half is.

So whatever we're doing on the front half to get the rendering, to get the clear delineated contents from water, continents from water, we've got to get down without affecting the front.

Can you put that PDF on the desktop so it's easier to find?

We're going into comfort mode.

We're actually getting pretty dry on viewers, so maybe it's time to end the stream here. I, uh, I'm enjoying working with you guys, but I think I want to go listen to some music now.