← Back to search

Built with AI: Day 3 with Jacob Posel and Cody Plofker

1 to 100 · 2026-11-03 · 35 min
relevance 59 8217 words Episode page ↗ Audio ↗
Show full episode description
In this episode of Built with AI, Jacob Posel and Cody Plofker join Jack to talk about AI-built landing pages and how they run their agents. Chase Mohseni joins briefly at the start. Cody offers brands a new landing page for $20K, paid only if it beats their current page in an agreed A/B test. If it loses, the brand keeps the page for free. Cody thinks a 5% conversion lift is a fair bar, and Jacob says that lift could be worth $50K to $100K to some brands. Jack shows a skill that copies a reference site or a whole store into a theme. Cody warns that the copied page may perform poorly. His own skill turns a reference page into a wireframe, asks about the product, persona, traffic source and angle, then builds in the brand's design system. He says brand context, such as heat map data and customer surveys, matters most. Jacob explains workers, a feature in the product he builds: specialists inside a general agent like Claude Code or Codex, each with its own skills and knowledge base. He says separate agents help people adopt AI and let teams give each agent different permissions. Cody has models review each other's work because agents grade their own work too easily. Jacob's Codex reviewer checks API values against warehouse data in a separate process with fresh context. He says agents improve when they can read logs, open the site and run tests. Jack keeps losing finished work on unmerged branches. Jacob suggests starting with a /ship command that merges and deploys, then adding checks when something breaks. Chapters
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Three operators compare how they actually use coding agents (Claude, Codex, Grok, Hermes, HQ) for landing pages, reviews, orchestration and shipping without losing work.
Benefits
  • Design system plus context layer lets a landing page one-shot to ~80%
  • Separate-model reviewers catch errors agents miss grading their own work
  • Hermes on a Mac mini handles orchestration while Claude/Codex build
  • HQ workers bundle skills, knowledge base and examples per vertical
  • A simple /ship command stops finished work dying on branches
Use cases
  • Cody's $20K landing page offer: paid only if the page beats the existing one in an A/B test, else free
  • Landing page skill wireframes a reference page, asks product/persona/traffic/angle, then builds in the brand's design system
  • Codex job ran 18 hours regenerating product images, posting results to Slack
  • Hermes on Mac mini via Wispr Flow captures ad ideas from the car into an ideas file
  • Claude orchestrates while Codex runs an adversarial data review in a fresh CLI process
KPIs / results
  • $20K performance-based landing page offer; ~5% lift considered fair
  • Landing page in 1-2 hours; one-shot gets ~80% there
  • Hermes on Claude API cost ~$30 in two days before switching to subscription
  • 28-person team, few technical, evaluating HQ
Tools / build
  • Cody's landing page skill
  • HQ workers and /ship command
  • Hermes agent on Mac mini
  • Codex adversarial data reviewer
  • Bake-off skill running Codex and Grok from Claude
0:00 / 0:00
📑 Chapters — tap a time to jump there
00:00
Trying new models
  • Kimi K3 waitlisted; OpenCode only has older Kimi models
  • Codex job running 18 hours regenerating product images to Slack
  • Inference.net as open-source model provider
02:42
Cody's $20K landing page offer
  • $20K if new lander beats existing page in A/B test, otherwise free
  • Debating what lift justifies $20K; ~5% fair, not 3x conversion
06:13
Copied pages, design systems and context
  • Replicating Momentous-style pages with GPT as baseline theme
  • Cody: design system and context layer matter most; knockoffs may be bad pages
  • Landing page skill: wireframe, four questions, one-shot build
12:12
Workers, skills and separate agents
  • HQ workers: own directory, config, knowledge base, examples, multiple skills
  • Separate agents mainly for UX and per-agent permission levels
  • /new worker and /run commands
17:21
Models reviewing each other's work
  • Run Codex and Grok from the Claude app; bake-off skill
  • Agents grade their own work too easily; fresh-context Codex reviewer
  • Claude Max usage burned by Fable fast mode; Grok 4.5 as fast sub-agent
20:44
Cody's Hermes agent on a Mac mini
  • Cody forked Hermes self-learning into Claude and Codex
  • Hermes on Mac mini for orchestration; Claude remote control can't do two devices
  • API cost ~$30 in two days; switched to subscription, first Codex moment
23:13
Long agent runs and self-checks
  • 4-8 hour runs regress with Opus and Fable; smart models hyper-focus
  • Snapchat creators misplaced in search results of influencer tool
  • Give agents verifiable checks: logs, open site, Codex computer use
26:55
Bringing a small team onto AI
  • 28-person team in Philippines and South America on Canva/Figma
  • HQ pitched as better memory system; tutorial and case studies
29:21
Shipping finished work with /ship
  • Work lost on unmerged worktrees across 8-10 parallel tasks
  • Start with bare /ship: merge and deploy, add complexity as things break
  • HQ marketplace engineering pack; policy to recommend shipping when done
I'm not sure if you guys have tried out this Kimi code thing. Everyone's talking about it. I tried to sign up and apparently there's a waitlist. So I haven't been able to get an account. I'm still old school using Claude and GPT. But I wanted to try out... I was using the last model. The pricing is incredible. And they even had like a two week span where like, essentially tokens were free so you can just run it up during that period. So it does what it needs to do. I haven't gotten into the new model though. Right, so the K3 model, right? Yeah, I like literally, I got open code to see if I could get access to K3. Because apparently open code has the Kimi models and it's all the old ones. And it seems like open codes for like people who just want to like exclusively use the old models. Let me see if I can show you guys this. I'm not sure if you guys are trying out open code yet. It's like, I don't know, multi model agent thing. And, uh, looks like I have to launch it again. So it starts on this like page. I don't know. Let's, uh, audit image gen or products. I started this thing last night. It's literally been running for like 18 hours on the like actual codex app, um, which is generating like product images. Because I had these shitty AI product images. They were really like, they were fine for like, I don't know, um, four months ago, but now they're shitty, right? Because AI images have got so much better. Um, and it's like shooting images over to me in Slack. So I don't know, let's audit it and see if it thinks that it sucks. So you're running here. And I actually, I stopped it while we're chatting right now. I'll show you it in Slack, but, uh, I'll get it to audit it and see if it can. Um, but yeah, it's actually like if I switch models, it's actually like all these free models. I'm not sure if you guys have tried any of these. And then, uh, it's got these, it's got like GPT's latest models and, you know, a few of the new ones, but it doesn't have Kimmy only has the like old Kimmy models. So I can't try it out. Have you guys had an opportunity to try out yet? Kimmy? Yeah. Like K3 or whatever the new one is that everyone's pumped about. I haven't used K3. Um, when I went to go try use it, it was blocked for consumers. I, I, a friend of mine has a company that runs open source model. They're an inference provider for open source models and they're meant to for it. Let me see, but I haven't tried it. Here we go. This company. Inference.net. Did you say imprints.net? Do you want to show your screen? Inference. Inference. Inference. Dude, I was, I was doing something else. I was doing something yesterday and it tried to off inference, uh, in a browser, not inference.net inference.fmch. Jack, we should talk about Cody's landing page bounty and how we productize it and the future of marketing as decentralized. This is actually pretty good for you, Chase. This, I think it's cool. And I think there's like a, this generalizes to all, um, creative assets and we should talk about Cody's bounty. And if he's doing it with AI and I'll be, he's actually going to analyze performance. So I'm interested in. I was going to. I know. No, I'm here. Are we live? Yeah. We're live. Oh, I can tell. Yeah. No, I'm just like fucking around. I'm like, what would, what would be fun? And I don't know. That's like where I've spent the most time is building like landers or doing zero, but I feel like that could be a pretty easy one. I mean, I can like been building, been building some landers this week for a new thing. And like, you can build a pretty good lander hour, maybe two hours. I'm trying to get that down. It's like that first, you can one shot it and it's like 80% of the way there, especially if you have a design system. And then, you know, if you're trying to perfect it, if you care about like how it looks, then it's like multiple hours on the next step. But otherwise you could like one shot it pretty quickly. What's the hard part? Like, how did you get those like crazy cool designs? I saw the ones that you made with like the squiggly lines and the cool breaks and stuff like that. Is that from the design system or is that your taste? I don't know which one you're talking about, but probably design system. Yeah. Maybe that was for like my personal one. It's just building the design system and component library. Like that's the thing with all this stuff in context. It's just like taking the time to do that. And I mean, I did it in a weekend, but just building, you know, a markdown file and then obviously turning it into like a component library and just giving it the context. I mean, that's what you should be selling people like build them the design system, build them the deal to do it and enable them. And so you don't have to do the delivery stuff. I'm doing some consulting for that, but I think you got to you still got to sell an outcome, especially because that's all just getting easy and commoditized. I don't know, like I'm doing a little consulting on it, but I'm like showing people how to do it. But I think I don't know. I think it's the outcome. It's like that's where, you know, the my for people like it was like a 20K offer just because I had seen some people do it. I'm like, why not take a few cracks at it just for fun? If you know, if it doesn't work and I don't it's it's so essentially offer is like if anyone wants a new lander, as long as I qualify them and have enough traffic. And if I think they can win, we'd have to agree upon what success would look like if it beats their existing page. I get 20K. If not, they get a free lander, you know, and so in my mind, it probably takes a few hours as long as the brand has the right context that they would be able to give me because I want like the right insights and the right data and stuff. And then and not just like wing it from scratch. But yeah, I think the I think the design system would be pretty easy to build now, especially how good the models are. How much? How do you measure performance? Like what is your lander winning actually mean? And is it different to different brands? It would have to be. Yeah, it would have to be a be tested and I would have to agree upon it with the friends. I was talking to a few people because I'll be reached out and some people are like, oh, yeah, you'd have to triple our conversion rate or I'm like, that's not gonna happen. Like, but it's kind of it's like what would what it what is a 20K lift worth, you know, or like what would make somebody happy to pay 20K? Is it a 2% conversion rate lift? Is it 5% part of it depends on traffic, like the brand needs enough traffic to test, but also like that a that a decent lift is good enough. So I don't know. I feel like 5% is pretty fair, but some people were asking me to like triple their conversion rate. I'm like, I'm not gonna I'm not gonna, you know, build a page up front for a 300 and only get paid if it's a 300% lift. I mean, there has to be some way that you can just determine this. Like you can just determine what is like objectively what is 20K worth to them or like at least probabilistically what is 20K worth to them and like align or some number. And even to some brands like it's got to be way more than 20K if you're getting like a 5% lift and conversion for some some of this traffic like surely that's worth 50K 100K to some brands. Yeah. Dude, I need you guys to tell me why this is a why this is gay and I think that you might be the perfect person Cody to call this out. Why it's why it's what why it's gay? Why it's gay? Yeah, why this why this is a bad idea. I'll show my screen. I can show you what I'm talking about. So I built a why I didn't build anything. I just asked Claude to do it for me. And actually GPT is doing a better job. Like initially like I guess start building the scale with Claude, but a GPT is smashing it. It does a way better job of getting accurate. But, uh, yeah, I can replicate landing pages from pretty much any website. Although the whole theme, the whole like store into a theme as like a baseline for those people who I don't know, want to reason from analogy. They say, hey, Momentus knows their stuff. Grunz knows their stuff. So I'll like use that as a baseline template. This is the opposite of what you were saying, Cody, right? Because this isn't a design system. You have to apply it in science system on top of this because you have a bunch of brand like stuffed into this, right? If you're replicating an existing brands website. But, uh, but yeah, it seems like it's pretty good. You know what? Like a moment has had this beautiful website. That's like, it's so nice. So I was like, okay, you know, um, I was selling supplements. Could this be a good baseline? Right. To work from. So I'm not sure about you guys, but for my brand, I got like a bunch of, I got called to make a bunch of landing pages from like scratch and they were okay. But it was, you know, kind of really biased on the information that I was giving it as like, hey, do this. Um, you know, I want a section with a slider that does this, or I want a section that looks like this. I want a section with this website. Um, whereas this is like baking in a lot of the genius that's in these teams that I don't know, momentous or, or grooms or whatever the target pages. I mean, you're also just assuming it's genius. Like I will see people knock off Jones road stuff and I'm like, that page is actually doing terribly. So I don't know. I just would be like really thoughtful about it and know that, you know, what you're doing and, and taking, but I don't know what it's like, what's the point if you're trying to knock off momentous, great. If you're trying to build a, you know, something for a new brand or whatever, like you gotta, you know, you really want to make sure that like you, it matches the design system. So that is so much of it. And I think the, I would like, I have a skill, I have a new landing page skill that I built and refined over time where like, it'll take a page, I'll give it a new page and it'll analyze it. And it'll essentially turn it into a wireframe. And then ask me like three or four questions, like what's, what's the product? What's the persona? What's the traffic source? What's it? What's the angle? And then it'll just go and make it. And that's where it like will one shot it. But obviously in whatever, in whatever design system, whatever company I'm working on with that design system. So like for this new thing, I'm building spun up three landers, you know, this week. And they're, I mean, they're, they're, they're built and trained on some of like the really good ones. So, you know, I think the, the designs, I'm at least for me, the design system and like that context layer is the most important part. Right. It has to like inform itself, right? Because when you make three landers, you probably don't know which one's going to work the best, right? But the one that works the best is probably going to inform, give me some data that informs, hey, what's the next landing page that we make? Yeah, I think that maybe that's a little bit more important than like time to product and reasoning from analogy is actually that competitive process of like, we made three landing pages, which one works best? What do we learn? And what's going into the next landing page that we make? Right. Well, no, I mean, yes, but I'm not even talking about that. I'm talking about like, if I'm, I'm just making it up. But if I'm, if I'm Groons, right, it's cool that you can take a momentous page and, and knock it off and create a new momentous site in a day. But like, if I'm Groons, that doesn't help me that much. What helps me more? Right. Maybe I shouldn't say the brands, whatever. Right. But if I'm Jones Road, and I see a Groons page I love, like, how do I apply those fundamentals and get that to be a Jones Road page that quickly? Right. Right. Right. If you have, whether it's an HQ, whether it's locally, whether it's in a repository, and you have all of your stuff together, and then you have skills that are built on it, or you have a design system skill and a landing page building skill, you know, design skills. You can then say like, hey, that's a really cool thing I saw. I love that navigation that company is doing build one for Jones Road, and it can even take qualitative and quantitative insights that it has. And it's like, all right, cool. Here are the things that we should merchandise based on. I have our heat map API revenue per click in there, or I have every survey the brands ever done. So I know what the main, you know, objections are. And so I just what I found is like, just dumping all of that stuff and have it be available by your by your agents. Like you don't even know how it'll be used, but you can just that's where then you can like do some of this like goal on codex or like loops because you just it has access to everything and then you have to give it away to like self verify. Cody, you should speak to the big labs and build them RL environments for marketing, creating marketing and landing pages. Who do you know, like anthropic open AI, they're all paying like eight figures. No, not six figures for experts who can the top ones are making seven figures, but for experts who can build them basically environments to help LLMs learn how to do tasks. So given you're probably one of the best in the world at building landing pages, design systems, conversion rate optimized pages using AI. You could probably build them a couple of RL environments. I mean, man, if you know anybody. Yeah, somebody when I went to somebody hit me up and they're like, hey, do you want to work at anthropic or whatever? It was just like like somebody not anthropic. And I thought I thought maybe he like knew somebody and he's just like, oh, I saw they're hiring. You should apply. I was like, I thought you're about to make an intro. Yes. If anyone from anthropic or open is listening. All right. Let's talk HQ, Jacob. I am getting frustrated because, you know, I'm using my mini and I don't know if it's like the login issue, but sometimes I try to do a handoff and go from, you know, my laptop to mini and it doesn't have access to it, even though I've done a handoff. Is that because it gets logged out? Do I have to run a sync? Because like usually it's just a one. It's just handoff and then resume on the other device. Why does that sometimes not work? Can you check if auto sync is turned on on the mini and does it is it logged into the same account? And as the mini hasn't gone through the full setup where it's got a fully configured cloud company on both device. It should because it works most of the time. It's just occasionally. Hmm. So sometimes like this morning, I submitted a bug. It was a login issue. And then sometimes it's a it's a sync. But it should auto sync, right? Yeah, no, it should auto sync. It's the same account. It's the same account. All right, then tell me about workers. I was a not was something about workers and then I was with Brett Curry earlier. He was something about workers. What are workers in HQ? Oh, cool. Well, let's let's sort out your sync issue. We'll get on. I think I figured I don't know. I got that working. No, we need no we need to make that work. So we'll get the depth put it in HQ there. But the workers are I think a novel ish HQ concept. Um, where inside of Florida codex. Um, you skills just sometimes aren't sufficient as as like primitives that you use for automation or for workflows, where a skill is meant to represent some automated task. Um, and then people think like, okay, now agent, like I'm going to go from skill to agent and agent is its own separate environment. It's own separate personality and a skill is just a file that I run on my computer. But even when I'm inside of code code, even when I'm inside of codex, I want my own, um, independent specialists, even within that environment. Like for example, I'm in code code. I want, I don't just want to see our skill. I want to see our worker and the CRO worker can have many different skills. It can have an analysis skill. It can have a build skill. It can have a review skill, but the worker should also have its own knowledge base. And it should also have its own artifact section. And it should also have its own examples. Um, so in HQ, when you create a worker, the worker is its own directory. It's got a configuration file, which tells it who it is. You almost like you're creating like your own little mini agent. Um, and then it can have anything you want inside of there. So for example, for the CRO worker, it can have examples of previous outputs. Um, it can have the design system directly embedded into the worker itself. And it's just a very clean way of managing context or specific vertical workflows or specific vertical automations beyond what a skill can give you. Cause if you just stuff everything into the skill, which is what a lot of people end up doing, or they start pointing towards different folders that are scattered around the computer within the skill, then it gets confused. So that's, um, that's how workers work. So, um, I would start thinking about that CRO worker and inside of HQ, there's a slash new worker command. And then you go slash run with the worker name to actually run the worker. I'm so bad at like, like even when I first heard skills or like even like took a while to get agents, like, I feel like I have to do things to, to, to get the concept of it. Like to me, that's like the same as like a skill or even potentially an agent. So I have to like play with it to say, so like what would be a use case for the worker versus an agent? Uh, worker is just, so the agent is Claude or the agent is Codex or the agent is Grog build. And the worker is just a specialized workflow within that agent. So it's just a surface where you're working, not necessarily the, the task being done. Some people will like, there's different ways of designing these systems. Like some people will have a general purpose agent and then they'll have, for example, our workers within that general purpose agent. But some people will have specialized verticalized agents. Like some people will just create a CRO agent or a, um, retention strategist agent or in chief of staff agent, for example. And in that case there is less need for a worker because it has a verticalized task and verticalized context anyway. But for the general purpose ones like Codecode, which is just inherently general, um, general purpose, you want ways to allow it to specialize. Yeah. I was gonna ask about that next. Like I have a Hermes and so he's obviously hooked up to HQ and like I've had him do a bunch of general things, but now I was going to go in there and start building some like more vertical agents, like a retention specialist or CRO one or something like that. Like when, when do you go and when can you use an existing more general agent with just contacts versus like building out an actual agent for that? I see all these people are like, Oh, I have, I have this agent and I have that agent. Like what's the benefit of each? In my opinion, a huge value on having separate agents is just UX. Like it's just the experience of being able to tag us. It's just like, I know I'm going to tag this agent for this task. And it's often you could just have a general purpose agent and you just give it one extra line in the prompt to say, now do the retention analyst workflow or something like that. But the experience of having different people, different personalities that you talk to for different tasks allow people to adopt AI a lot easier. That's the main one. And then the other one, which is less relevant to you, but relevant to our bigger teams is that with HQ agents, for example, you can configure different permission levels for different agents. So certain teams want an agent that only has access to these API keys in this context. But other teams don't mind just having an agent that can do everything. So that's another reason why they want multiple different agents. Yeah, that's a good point. Dude, I want to see the UI for HQ because if it's a, as long as it's better than Claude, then I'm down. What do you guys think about the Claude UI and how it's, do you guys use Claude? Like as in the app or do you use it in terminal? I mean, I guess Jacob, you're using it in, in HQ probably exclusively, right? Well, I'm going to disappoint you use HQ inside of Claude. So the app is cool. Oh, of course. Yeah. Yeah. You don't move. You don't change to a new surface. But yeah, I just, I can't get on board with this UI like home. Oh, really? I mean, these different Claude. I'm all of code. I like, I like, I don't know. I like Codex a lot, especially mobile. Like Codex just totally has it beat. One really cool tip that Jacob showed me that I did not even realize is that you can run multiple models from a Claude app. And so that's been pretty sick. And so I've been running Codex from that as well. And Grok because, because if you have a Twitter subscription. So I built a skill like a bake off skill. And so sometimes just to test and like sub it out. But the other times I'll like run a workflow. Like I, I was doing something Shopify and couldn't figure out a bug and had, had, I think I was in Claude at the time, but farmed it out to Codex and Grok. And like agents are like terrible at grading their own work. They're just like really easy on themselves. So like subbing it out and creating a workflow where you have different models or agents grading each other's work. It's like sometimes when you get stuck, it's been like a really, really good little, little tip. So hats off to Jacob for that. I was getting cooked. I was doing some analytics work for a company and I was getting cooked because the board wasn't coming up with good answers. So I built the most robust Codex reviewer, like data reviewer that will like back to the API values and then check the data warehouse values and then run an adversarial review against the data. And then make sure that the data interpretation is right. Anyway, super robust. Maybe I'll share that with you guys, but I run that within Claude. So I let Claude orchestrate and then Codex will actually run the review in a NS through the CLI in a separate sub agent in a separate process so that it's completely fresh context. And there's no bias from the session that you're in with Claude. That's smart too. Also you hand off. So you create a new shell and you hand off or you do it within the like you call Codex in the CLI. Like how you guys actually like functionally doing this? I see some people who would do it like they're working primarily in the Claude app or the Codex app. And like I'm kind of I'm back and forth with the Codex app because the remote feature is pretty rad. You know, being able to control the machines via the phone. But most most of everything I do is in CLI. How are you guys like actually calling in another agent, calling in Codex, calling in Grock to do an audit or like double check something? Is that all in the context of CLI as well? Are you all in shells? No, you can do the app. That's what I didn't know that Jacob told me. I think I just then once he told me that I just went to my Claude and was like, hey, I found out how you can do this. How do you do it? And I think it calls the CLI. Yeah, just tell it from the app. Just tell it to do it. Yeah, I did not know that. I had not seen anyone talk about that. This has happened to you guys as well recently. I'm on like a max. What are the max max plan is of Claude? And I haven't been using Claude at all because I've burnt through all my usage like, I don't know, a week ago. Has this happened to you guys as well? Like I've pretty much been exclusively, you know, GPT and Grock because my Claude usage has been down for days. Oh yeah, it happens sometimes. Especially if you're using Fable. Fable. Yeah. Fable, fast mode, just like blasted through my credits. Yeah. I think I got like your license or something. But yeah. Great user experience. Use Grock or five as your sub agents with Claude. It's speeds things up significantly. If you do like longer running tasks, Grock for five is super fast and you can invoke it by CLI the same way that you do with codex. Oh, nice. Um, yeah, I need to try that out. I need to try out Grock more. I didn't know you could get a plan through your, your like Twitter plan or whatever your ex plan. Yeah. But, um, but I think I hit a few prompts and it said, hey, you ran out, sign up and I don't know. Did you guys hear about the Grock uploading everyone's repos thing? Is that like, yeah, I saw that. I don't even know what the stage were open source. So not the worst thing in the world. Right. Right. On the topic of open source. I wanted to ask you, Cody, you said that you had set up a Hermes agent. Um, Hermes is open source, right? Um, I did some, uh, something funky cause I kind of wanted what Hermes provides, but I didn't want to run everything through API. Um, I think that Hermes when it was first getting a lot of hype, you could only use it with API. You can use it with your accounts. Um, so I just like added the self. I just forked the Hermes, uh, Hermes into Claude and into codex. It's like the self learning, getting better, um, tool. But then of course I'm leaving behind whatever new updates they, they make to Hermes. Um, are you guys getting a lot of value out of it? I love mine. I, I only, so I got a mini cause I wanted, I was doing a lot of Claude remote control and there were times I was doing a lot of Claude. Sometimes where like my laptop is closed. If I'm like going into New York or something like that, that I want to keep going. I found out by having that, that it's not that easy or you can't really have two devices on, on a Claude remote control and you can't really start a new session. So I was like, all right, let me, let me get a Hermes while I have the mini. Right. So I can use them more. So I set up a bunch of stuff to sync them. And I, I wasn't sure like every, anything that somebody showed me on their open claw. I'm like, well, I can do that in Claude code. Like I don't necessarily get it. Um, but it's been really powerful. Like, and not instead of like, I do all my coding and stuff like that. And, and codex are in Claude, but like, I just do a lot of like orchestration and even like, like I was in the car today, listening to a YouTube video and I'll just like yap and whisper flow like an idea. Like I had an idea. And so I have like an ideas file and you know, one of the company's, um, HQs, I'm just like, Hey, save this ad idea. I think this would be a great idea. Um, so it's, it's like, for me, it's less of the builder. It's more of like the orchestrator. We're like all Claude or codex and build stuff. But so I had it running on Claude API for the first two days and I was spending like 30 bucks and I'm like, this is unsustainable. And I found out somebody on Twitter told me that you can just run it in subscription. So that was the first time I actually got like an open AI pro plan. And then I was like, well, I have it. Let me just start playing with codex. And then it was my, my first codex goal moment was just like mind blowing. And I was like all in on Claude before then. And I don't know now, especially with cheap with 5.6, like, I don't know. Um, they both have value, but, um, if I had to pick one, I'd probably go codex. Oh, a hundred percent. I'm thinking of churning off a Claude. I haven't been using it for like four days. Um, while my usage has been out and, uh, I feel like every time I switched to Claude, it just fucks things up and it takes so long to fix for codex. So, uh, yeah, I don't know. Um, yeah, I'm seriously considering churning off of, uh, Claude. And if like Kimmy is going to get good and Grok's apparently good, then, you know, I've got other options. I don't need to, um, you know, have Claude for like two days and then it's the weekly allowance or the, you know, monthly usage is, is all done. Right. And, and my projects are all fucked up until they release until they release. And then my projects are all fucked up. I'm not, I'm not sure if this is, uh, you know, I have like a certain hit rate with, um, uh, my projects, right? Like with whether we actually achieved the thing that I wanted to get done. And the, I guess the problem with this is that sometimes they're running for four hours or like eight hours. And then after we haven't made progress, like it's actually gone in the wrong direction. There's been like a bunch of regressions. Um, and I'm seeing that a lot recently with Opus and with Fable. So it's a, yeah, it's a weird one, you know, I'm running out of usage. It's made, it done a load of damage. And then I'm like, okay, do I actually need Claude? If I can just use different models that I'm going to destroy everything and then leave, you know? Yeah. Uh, five, six does that as well. Sometimes though, um, it's a smart model thing. They get like too hyper-focused on some issues sometimes and unable to progress. It seems to be a problem with smart models. Oh, I was going to say, it seems like a skill issue. I mean, Hey, is it, there's a, it happens. Yeah. But you know, there's a hit rate to everything, right? Like sometimes things go well. Sometimes things go poorly. I'm not sure about you guys, but, uh, I'm very like hands off the steering wheel. And maybe this is the skill issue, right? After a certain point, I'm not going to provide any more context. Just like go do the thing, right? Use best judgment, bypass permissions, YOLO mode. Right. Um, but Hey, you know, I've got, I've got a little bit of freedom, uh, to do that. And then when it, you know, when it messes up, that's, uh, you know, um, I guess the non-desirable outcome, but the desirable outcome is you have, you know, eight, nine agents running simultaneously and they actually getting shit done. Right. Um, it's just, yeah, it's a little bit hit and miss, you know, every now and then, you know, it goes the complete wrong direction. Uh, and you would be surprised, right? Like I'm building this thing right now and it has a search function, right? For, for, uh, I can, I can say what it does. It sources influences across meta, um, Snapchat, YouTube and Tik TOK, right? Lots of brands are doing Tik TOK shops right now. Um, but not a lot of brands are doing meta because it's actually like really laborious time costly to reach out to creators on meta and get them onboarded and process them, right? To receive the content, to, to, to manage the payouts. So you need a team managing it or like at least a few people, um, you know, who have, who have processes in place. Um, and yeah, so like it goes off, it sets up the whole Snapchat integration. We've got full like API access through Snapchat. Snapchat are great. Yeah. They're super easy to work with, but, um, and then it places the, all of the Snapchat creators above the results for the search, right? There's a clear table for where the, all of the creators are supposed to pop up with all of their data, right? And the, you know, how much followers they have, what other platforms they're on. And, uh, yeah, it just puts all of the Snapchat creators in a separate card, right? Instead of in the results. Yeah. It's just, uh, I'm not sure if this is your experience too. Sometimes it just misses, you know, in some ways and, you know, in the, in the process breaks everything. Right. Um, but that, that wasn't, that wasn't the biggest mistake. That was easy to fix. It was just like, you know how this is obviously where the, the creator results are supposed to be. Right. Yeah. Should have like Claude valid or codex validate its work there, but open up a computer, have a look. And then it should hopefully loop around and figure out that it's wrong. Were you guys jamming on this sort of stuff back in like August, September, October last year? Um, cause I remember like, cause I'm, I'm totally non-technical. I'm not sure about you guys copying logs out of railway into cursor to like fix things and smoke testing manually myself, like actually opening up the app and checking it every time something was deployed. Um, yeah, the, the fact that we can, we can, the app can smoke itself now is, is awesome. Um, yeah. Yeah. It makes a huge difference. It's a huge difference. Like as much as you can to give an agent some verifiable way to evaluate its own work, um, changes things significantly. Like let it look at logs, let it open the website, let it run testing locally or even with compute, with codex computer use. It's so good now. Like there's just a lot of ways that, um, you can really improve, improve things. Do you have the time, uh, to show me HQ to give you context on my team? I have like just a couple of people in the Philippines and in South America. Um, a couple of devs, the technical and they think I'm stupid and a couple are like a lot are design based, right? So they do like email, do like graphics and stuff. And they do the product images on our website until recently, you know, I'm, I'm changing up the way that we do our product images. Would HQ be a good fit for, for us? I mean, there's like, you know, 28 people on the team. Um, only a few technical people, but, uh, I really want everyone to start using AI more, right. And stop using, um, you, you, you know, Canva, they're using Canva for design stuff and figure off a design stuff. And it, it's not a problem because, you know, we've got like a bank of good resources and, and a good workflow is pretty efficient. Um, the way that they're working in these design tools, but like the, the real problem is that they're not using AI as much as I would like them to be using it. Um, and it's kind of hard to say, Hey, you know, open up a CLI download, uh, codecs and just figure it out. Um, so yeah, is, is HQ a good fit for us or is it more for like, you know, enterprise? No, I think it is. I think it's even good for like small, but individual people. Like it's, it's a better memory system with AI than just using cord or codecs by itself. So yeah, I would say even for you to just try it out and use it and then, um, use the free one. Um, I, I have to leave in eight minutes. So I, um, I don't have time to onboarding would take more than eight minutes, but I would say for sure, just use it for yourself. It's probably a good fit for your team. Um, I'm actually curious cause I just released a new desktop app. Um, so I'm curious to see how it goes, but is it easy to self serve? Um, people find it incredibly easy. Some people struggle. It like, I think it just depends on how much AI you've been exposed to. So, um, I think you'll be fine. I think, you know, if, if some people can figure it out, I think I can, I can figure it out. Um, I mean, we also have a tutorial, we have case studies, so you can follow the tutorial if you, if you get lost. You know, I imagine it's one of those cases like as old as time where, uh, the people who struggle are the people who fly past the tutorial and ignore it and they didn't listen to anything and they go, fuck, how do I use this thing? What, you know, how is this different from, I don't know, just codex or, you know, call off a team. Right. Try it. See how it, see how it goes. Great. Yeah. I'll give it a go. I'll give it a go. Are there any things that you think would be, ah, this is a, this is a good question for you because dude, I'm struggling with this. I've spent so long when I say I'm struggling with it. It's because codex and cord couldn't help me solve it. When I'm chatting to him, like fix this. And then it just builds a convoluted, like, um, release governance, like process. But, um, basically I'm not sure how, how you are when, uh, when you approach building for HQ or building for any of the clients of HQ, the businesses that you're working with. But I'm very additive, right? Like when I build something new, I want it to sit on the shoulders of everything else. I don't want dirty work trees, branches, all this nonsense that I have to review the diffs. I'm not reviewing any fucking diffs. Right. I just want it to go live. Right. I guess like the, the simplistic, like the simple answer would be like merge and deploy like automatically. Right. You don't want to deploy every, every time you make a change. Right. But, uh, if you know, this is what I was doing before. I was merging and deploying pretty much everything. Um, and it was working. It was fine. Like nothing was getting left behind. But recently, you know, especially when I have like, I don't know, eight, 10 things going at the same time, some stuff's getting left behind. Right. Um, and that's a, that's a headache, especially if I spent a long time on it. I'm like, fuck, you know, I didn't realize, but this whole thing that I was working on for like two hours isn't in the app. Right. Um, it's on some work tree that never got merged that never got deployed. Is this something that you've struggled with, found a solution for, um, seen with anyone else? Cause like, man, it's the bane of my existence. And like, like I said, I tried to fix it with codex. I was like codex, any branches just fucking merge them. Like I want everything to go live. Right. I didn't work on anything that I don't want to actually go live. Right. Like, cause I worked on it. Um, but, uh, then it just built this really like convoluted, long release governance, testing QA cycle where it smoke tests. And that was like, it was just beefy. It was making like really quick things take like, I don't know, 20 minutes, 10 minutes, uh, to, to go live and, you know, longer things take a lot longer. So, so yeah. Have you got any thoughts on, on that problem and, um, any solutions for it? I think a bit of it is like just, um, building discipline about if you work on something, just ship it. Like don't, don't let it hang. Like if it's finished and ship it. Um, I mean, I would just create a slash ship command. That's what we have as does go through that whole, like, UA testing, smoke testing, like everything process monitoring, like put them crazy monitoring loop for like, you, I was afterwards to monitor to check if anything's broken. But, um, just like create a super simple one, like slash ship. But just merges the branch in and then deploys your site. And then just add any new complexity you need on top of that. And just build a discipline around. Flash ship. Or if you were to use HQ, you would probably create a policy that said, whenever it seems like I'm done with the project recommend I ship it. And then it would just ship it for you. Uh, that's the best way to do it. But there's no like really easy automatic way to ensure that, um, everything is always going to be merged into main because like by design, it's not meant to be everything to be merged into main by design. Otherwise branches wouldn't exist. Um, or you could just always work on main like levels, like Peter levels does just always work in prod, always work on main. And then just, um, just all changes are all live automatically deployed because you're just working on the VPS that your app is running on. Yeah. Hey, good shot, huh? Cause, uh, it's not, it's not that everything is ready to ship on every change. It's when it's finished. So you just need a reminder that, Hey, this is finished now ship it as opposed to this is finished now on this branch, uh, in staging. Um, yeah, it needs to go live. Right. Um, I'm not sure about you, but like my entire process is if I start on something, we're going to finish it. Right. Even if it doesn't work the first time I'm going to hand off, I'm going to, you know, see if Claude can hack it. Um, uh, or vice versa. Right. I'm not, I'm not building anything with the purpose of like leaving on a branch, but that is the distinction here. Not everything is done. Right. Um, you could be, you know, I don't know, three prompts in versus, you know, maybe it takes you 10 prompts to get to the finish line and, you know, twice as long. Um, that's the only point where you want to merge it. Right. Is when it's actually done as opposed to at every milestone, which is what I was doing before. Right. Everything was always merging to staging. Right. And then like, you know, uh, to production when it was, uh, actually ready as opposed to existing on all of these different branches and then going, fuck, where's that work that, uh, we spent an hour on. Um, right. So did you say that the slash ship command in HQ, um, that just ships it or does that do that whole process where it QA is it for you as well? And you know, it does that thing, but you could, you could build your own. It does. It does the whole thing, but you could build your own. Um, just like tear it down to the bare minimum, like merge and deploy and then new things as you need them. Like the reason why we need that is because, um, we have a lot of users at this point. And so we want to make sure that we want to be certain about things when they land. But, um, we also have a lot of microset. We also have a lot of services. We're not just a monorepo. So there's complexity around merging that we need to take care of. But, um, like I would say just start with the bare minimum of what you need and add complexity as things break. That's a good shout out here. Yeah, that's a good shout out. Well, that's ship. And if you download HQ, you go shift. Mon O sorry, shift option. O you go to our marketplace. Share it with you. Go to the bottom. You can see our, um, engineering pack. Um, that's where you can install all of our different engineering commands. Yeah. Good shout on, um, you know, when it breaks, then fix it because, uh, yeah, that's, that's kind of what I'm in the process of doing right now. It's like, Hey, why are we leaving work behind? Um, you know, we worked on this for a reason. Let's get it live. Yeah. I'll check that out. And, uh, your marketplace is basically a, um, a collection of skills and workers and stuff. Yeah. It's just a collection of packages kind of like NPM packages. Um, that you can just anyone can install into the HQ. Oh yeah, dude. Nice. All right. You said that you have to go. I want to make you late for your next call. Catch you in the bit, dude. Thanks so much. It's good to see you. You too, man. Bye.