Riley Brown Automates Video Packaging With Codex to Shun Mass Batching

Riley Brown借助本地Codex实现调研自动化，拒绝批量化视频制作 大规模内容批量化会产出缺乏灵魂的媒介并削弱创作者壁垒，这意味着AI智能体应被用于压缩调研流程和提升制作工艺，而非单纯追求产量自动化。 Riley Brown拆解了他是如何将Codex等本地AI环境与专用工具相结合，在不陷入机械化批量生产的前提下，运营其拥有170万粉丝的内容与咨询业务。他展示了具体的工作流，包括渲染程序化Remotion动态图形、使用Supadata抓取爆款视频逐字稿并通过BRENS框架构建视频钩子，以及在Paper中快速迭代缩略图。Riley还阐述了自己的理念：内容包装应次于深层且真实的好奇心，通过Slack智能体实现团队运营自动化，并始终扎根于现实世界的软件咨询业务，以此确保其教学内容保持真正的价值。 Riley Brown在其拥有170万粉丝的业务中，通过Codex进行本地AI运行来自动化处理调研、动画B-roll及缩略图重混，同时严格避免使用AI剪辑视频和批量生产，以保护制作工艺。 (0:00) 通过在Codex内部运行Supadata API子智能体，Riley可以在约30秒内为整套YouTube创作者视频库建立索引，从而反向推导头部频道成熟有效的30秒视频钩子结构。 (7:28) Riley将Codex与AI原生画布Paper相结合，抓取竞品的爆款缩略图，并通过Nano Banana执行视觉提示词换脸、光影调整以及快速生成用于A/B测试的变体。 (25:21) Riley并未单纯依赖媒体赞助，而是将业务立足于企业AI咨询及其初创公司Chorus；他认为，在实战一线部署真实智能体，才能避免内容创作陷入枯竭。 (38:16)

Riley Brown argues that the lasting competitive advantage for digital creators lies in high-craft quality over automated batch production, using modular desktop agents to multiply research depth while grounding content in active software development. The Quality Moat and Deep Research (0:00) When creators first adopt AI automation tools, their default instinct is often to churn out high-volume batches of scripts and posts. Riley Brown rejects this model, warning that batching inevitably turns content soulless over time. Instead, he treats agentic environments like Codex as always-on research assistants that operate asynchronously in the background. While drafting, he tasks agents with fact-checking claims, fetching source documentation, and analyzing the retention mechanics of top creators for five to ten minutes at a time, effectively accelerating research density by 10x. This research leverage allows Brown to target an aggressive cadence of four long-form videos and five short-form videos per week without burning out. Rather than using AI to generate full teleprompter scripts—which he warns erodes audience trust as artificial content proliferates—he scripts only the opening hook and delivers the remainder off-the-cuff from direct technical experience. Modular Skills and Visual Storyboarding (3:16) Rather than relying on isolated chat prompts, Brown structures Codex into composable "skills" that chain distinct capabilities together. To automate video intros and B-roll, an agent analyzes draft scripts, calls SerpAPI's Google Images endpoint to pull high-resolution software logos, and feeds those assets directly into Remotion code to export branded motion graphics matching his visual style guide. When maintaining these tools, Brown never manually edits underlying code; he provides natural-language feedback on errors, instructing the model to revise its own instructions before verifying them in fresh sessions. Brown also bypasses manual scriptwriting by dictating unstructured thoughts into Wispr Flow using AirPods Max while pacing around his room. A custom Excalidraw skill transforms this continuous brain dump into a nine-slide diagram deck that establishes roughly 80% of the final video's visual and structural pacing, anchoring the production visually before he spends 20 to 30 minutes fine-tuning diagrams and filming demos. Curiosity-Driven Packaging and the BRENS Framework (13:24) Brown pushes back against conventional advice to finalize titles and thumbnails before shooting, arguing that letting packaging dictate creation drains creator energy. Instead, 90% of his effort goes into following personal curiosity and documenting unscripted technical tangents. After filming, he backwards-engineers the packaging: the raw transcript is fed to an agent running a custom Supadata-powered YouTube research skill, which scrapes high-performing channel formats in seconds and designs hooks tailored to what was actually covered. For long-form retention, he applies the BRENS framework—making hooks Big, Relatable, Easy, New, and Safe, reassuring viewers that investing 30 or more minutes will not waste their time. For visual packaging, Brown uses Paper, an AI-native design tool programmatically linked to Codex. The system scrapes top thumbnail layouts from creators like Alex Hormozi and Andrei Jikh onto a canvas board, executes in-painting face swaps to insert Brown's likeness, and generates batch variants in parallel. By pointing the model directly to visual references on the canvas to match typography and rim lighting, he skips complex written style prompts, handing high-fidelity concepts to a human designer for final polish and continuous A/B testing. Multiplayer Agents and Real-World Product Building (33:55) To coordinate automations beyond a single desktop, Brown utilizes cloud-based agents hosted on his platform, Chorus, integrated directly into Slack. Rather than building a centralized agent that attempts to manage every business domain, Chorus deploys specialist personas—such as dedicated CMO, developer, and content ideation bots—that converse with team members in shared channels. Brown notes that automation currently involves operational friction: local desktop agents require hardware like a headless Mac mini to stay online 24/7, while cloud agents struggle with persistent authentication and local file access, though he anticipates advancements in cloud browser-use models will resolve automated login workflows within the year. Addressing creator monetization, Brown argues that relying purely on sponsorships, ad revenue, and corporate training turns creators into superficial news aggregators. By actively building software products like Chorus and Vibecode alongside engineers, he learns multiple technical lessons daily. This frontline engineering work supplies the practical substance and failure-tested lessons required to produce credible, differentiated educational content. Brown emphasizes that despite heavy agentic automation across ideation, graphics, and research, long-form video editing remains entirely human-driven because AI video editing capabilities are not yet sufficient. Furthermore, scaling agents across teams remains constrained by unresolved trade-offs between local file access, continuous machine uptime, and cloud session authentication.

[0:00] Riley Brown: What I want is better, because I think the moat is quality over a long period of time versus batching. Batching will make it soulless over time. [13:24] Riley Brown: By the way, this guy's an expert, he's going to talk about these things, and I promise you by the end of the video you will get value out of it. [15:09] Riley Brown: I feel it in my heart, like my gut tells me what I should talk about [15:09] Riley Brown: Then I'll end up making dance videos on TikTok, which is something I don't want to do. [17:24] Riley Brown: turn this into a skill. [17:24] Riley Brown: Hey, I want to make a short-form video on the latest Meta Ray-Bans. Please add a new short form in progress to Notion and give me the link. [25:41] Riley Brown: Please replace the guy with a beard with the guy that doesn't have a beard, just replace the guy, except make his shirt white, and make him wearing a chain. [31:15] Riley Brown: Hey, you didn't do a good job, please change the skill so that you don't do that again. [32:28] Riley Brown: Just realize that there's no risk, just let your AI cook and tell it to do a better job. [33:55] Riley Brown: mini god that knows everything [38:16] Riley Brown: ride off into the sunset and make a very good living

[0:00] I will just scrape a hundred thumbnails  that performed really well and I'll [0:03] put my face on it. I took this and  then I converted it into one for me, [0:07] and here are my thumbnails. Turn this [music]  into a skill. The more skills you use, the more [0:12] you'll realize that you can mix and match.  That's where everyone's brain goes at first, [0:15] is to like batch and create more. What I want  is better, because I think the moat is quality [0:21] over a long period of time versus batching.  Batching will make it soulless over time. [0:28] Hey everyone, my guest today is the one and only  Riley Brown. Riley makes AI education content for [0:33] over 1.5 million followers. So super excited to  get him to show us how he uses Codex and other AI [0:39] tools to run his entire creative business and  more. So welcome, sir. Thanks for having me. [0:43] This is this is going to be a lot of fun. Yeah,  you know, I feel like you're pretty much on all [0:47] the platforms, YouTube, X, Instagram, and so on so  forth, like can you talk about your business at a [0:52] high level and how do you use AI tools at a high  level to kind of like create all this content? [0:57] Yeah, so at the center of everything that I do  personally, uh, I like to educate people on AI, [1:03] and I've been doing this for the past  three years. That's how I've gained almost, [1:07] uh, I'm almost at like 1.7 million followers  across all the different platforms. And obviously, [1:13] like, we've, uh, released a couple products, um,  including Vibecode and now Chorus. Um, and then [1:19] we're also doing, uh, a lot of like consulting,  uh, helping companies implement agents in their [1:23] business. And it's just been so fun over the  past three months talking to people, seeing how, [1:30] um, people are using AI agents in their business.  And you know me, I'm a content creator, that's [1:36] like my profession, and I use it for almost  every part of, uh, of my content business, [1:45] except for, like, video editing. I hire a lot  of video editors, and video editing isn't there [1:53] yet. But like for all of the ideation, planning,  coming up with good hooks, um, YouTube thumbnails, [2:01] um, doing all of my research, it's all grounded  in YouTube. So I have like a YouTube researcher [2:06] skill that my agent will use whenever I ask it  to do research. Um, and so like I have a list [2:12] of a ton of skills that just make creating  content at scale a lot easier using agents. [2:18] Got it. So so YouTube's kind of like the hub  of your content, and then like all the other [2:21] platforms you kind of make clips and share  social posts and so on and so forth. Um, [2:25] I make dedicated content on Instagram and TikTok.  That's actually how I got my start. I just prefer [2:31] longer form content, and I think Twitter  is a great place to post long form content [2:35] as well. My long form content does as well on  Twitter as it does on YouTube. Um, and so yeah, [2:40] I like making content native to all the different  platforms. Uh, clipping is actually one thing [2:46] I'm just now getting into, especially with AI,  but we can talk about that a little bit later. [2:50] Okay. So so can you actually show us how you do  this end to end? Uh, you don't have to make the [2:56] content, but like I would love to see the skills  and workflows that you have. Absolutely. And I [3:01] guess what I'll start with here is kind of how I  think about Codex in general, and platforms like [3:10] Codex. So right here we have Codex open. Normally  when you come to Codex, it looks like this, [3:16] right? It looks like ChatGPT. Um, but when you  go to ChatGPT, everything that you upload to it, [3:22] if you were to upload an image, many people have  done that, uploaded images or PDFs, etc, all of [3:26] that's stored in the cloud. When you use Codex,  all of it's stored locally. And so Codex, just [3:32] like Claude Code, can fully control your computer,  and in fact it has a computer use skill built [3:39] into the platform. It can actually control your  computer for you, which is really cool. And one [3:45] really useful thing that I do for my videos, if  you watch my videos, you'll see at the beginning [3:50] there has like really high quality animations. And  the way I do all of my planning for this is I use [3:56] this Remotion plugin. And so this Remotion plugin,  I actually have it open right here. This was one [4:02] of my best performing videos on Codex. This was  the opening frame. It was me saying, like, "Today [4:06] we're talking about Codex and all of the different  things that it can do." This graphic right here, [4:13] um, can you see it? Yeah, yeah. So this graphic  right here is a graphic that I just generated, [4:20] and you can very easily export this graphic,  and then I just placed it on top of my video [4:25] when editing it. And all you need to do to do  this, you can, this isn't my skill, this is a [4:31] built-in plugin into Codex. And whenever I want to  use it, I can just type in @remotion. Um, I also [4:39] have a Remotion best practices skill, which, uh,  it has like my creator brand kind of infused into [4:46] it. So whatever it generates will be in my own  like style, if that makes sense. But yeah, colors, [4:53] the colors and stuff. And so that's basically  how I use it for creating these little graphics. [4:59] So to make this graphic, you just, hey @remotion,  make the flowchart Codex logo, and then do you [5:05] have to provide the assets, or it just goes  off and finds the stuff? So this is another [5:10] thing that, um, I've set up. So I have a skill,  I actually, I think it's called, um, image pull [5:22] um, internet image puller. There it is. So this  internet image puller is one where it'll go off, [5:28] it'll actually use something called the  SerpAPI. Um, and so it uses Google Images [5:34] within the SerpAPI, and it will actually find  the relevant logos related to whatever video [5:39] it is. So if I say, "Hey, I need some B-roll for  this," it'll automatically look at my transcript [5:45] of the video that it's creating B-roll for,  and it will find the logos and it will just [5:50] pull them. And you can use multiple skills for a  prompt. So I would probably say something like, [5:56] uh, please pull the relevant logos and then make  a graphic for this, and then I would just use the [6:02] Remotion skill, um, which I call Remotion best  practices, and it will just generate it and pull [6:09] the relevant logos. Um, but you can manually just  drop them in as well. Okay, so you're basically [6:15] chaining multiple skills together, right? Is  that kind of, yes, I guess you could say that. [6:20] This episode is brought to you by Google AI  Studio, the easiest and fastest way to build [6:25] and share your idea with real users. For  example, I use AI Studio to create three [6:31] different landing pages for this bakery using  images generated by Google's built-in Nano [6:37] Banana model. One highlights fresh bread, another  highlights awesome cakes, and the last highlights [6:43] the bakery story. I can publish the website  easily and get a custom URL. In this case, [6:48] Corgi Confections dot AI Studio, to share all  three variations with potential customers for [6:54] feedback in one tap. When I was a product manager,  Google AI Studio was my go-to tool for building [7:00] prototypes like this. And it really is an easy  and free way to share ideas with your team, [7:07] customers, and everyone else. Check it out at  aistudio.google.com. Now back to our episode. [7:14] Let's go to the beginning, actually, like, we  actually still brainstorm ideas for the video that [7:18] you can make next. Do you do like a competitive  analysis, or like some sort of, what are people [7:21] talking about, skill? Yeah, so one thing that I  like to do, so my favorite skill is this YouTube [7:28] researcher skill, if you want to set this up.  Um, and I also, like, have a link to a lot of [7:33] the skills that I've created on my website, if you  guys want to check that out. It's free, I don't [7:37] charge for that. Um, great. But it uses this thing  called Supadata, which is one of my favorite APIs. [7:43] It allows you to pull the full transcript  from your videos, from any YouTube video, [7:49] like in one second. So I can scrape an entire  channel if I wanted to. I could just say, like, [7:54] "YouTube researcher, create a Peter Yang bot,"  um, and or skill, and it can scrape all of your [8:01] videos and it can act like you or something like  that. Um, and so this is my YouTube researcher [8:05] skill. And oftentimes when I use it, for me,  when I create content, the intro is the most [8:11] important part, and then the rest of the video is  kind of free flowing. I'm naturally good at kind [8:16] of explaining how to do things once the video,  once I'm into the video, but the intro is really [8:22] important. So another prompt that I have is called  a hook outline. So hook outline will create a very [8:28] specific type of document for me. And so I'm going  to say, like, um, please, uh, look at the latest [8:36] Alex Hormozi video and write a hook outline for  it. And, like, whenever I find a good YouTube [8:45] video, you can condense it down to a hook outline.  What is the intro, and then, what, and then just [8:50] outline the rest of the video. And what it'll do  is it'll allow you to, like, you can take that [8:54] exact format and apply it to a different video.  And so that's what I do a lot, is I'll create [8:59] a hook outline based on someone else's really  well-performing video, and then I'll tell AI to [9:04] take all of my ideas for a new video and convert  it into that structure, if that makes sense. [9:10] Oh, that's interesting, yeah, I'm following  it, yeah. So I've been using yt-dlp, but you're [9:15] saying that this other thing can just pull the  transcript in one second? Yeah, I mean it's just [9:20] like a quick API call, like yt-dlp, if I'm not  mistaken, it'll download the full video and then [9:28] pull the transcript, that is super data-intensive.  And, yeah, Supadata will just, you know, you can [9:36] say spin up sub-agents, that's one thing you can  do with Codex, and they have a really cool UI. [9:42] You can just ask it to use sub-agents, and it'll  split off into like six agents, and it will scrape [9:50] it all in like 30 seconds. You can scrape an  entire channel in like 30 seconds. [laughter] Got [9:56] it. Not, I'm not promoting that you do this,  but it is really good for research. So yeah, [10:01] I mean it is a fact that the first 30 seconds,  if it's boring, people will check out, right? [10:05] So basically you don't do a full voiceover  guide when you make a video, you just have [10:10] the intro scripted out and then you have like an  outline for the rest, is that right? Yes, exactly. [10:14] Yeah, I mean, I think with AI the  fully scripted videos are going to [10:23] go down in value. I think what will go up in value  over time with AI is you have to hook the viewer, [10:30] you have to play the game to some degree, but  you also just want to get good at talking about [10:34] the thing that you're super passionate about. And  so that's what I like to do, I like to show the [10:39] human side, and then having guests on is like  a really good way to do it, like your podcast, [10:42] you have a structured intro and then you just  talk to someone for the rest of the time, [10:45] and that's, I think, why people are here. And  so you can see here, this is a very specific, [10:51] um, format. Um, yeah, and one thing that's really  annoying about Codex, and one of the reasons why [10:58] I've been moving off of it, is because the GPT  models won't pull the transcript. You can see here [11:06] that it summarized his intro, and the reason it  did that is because of copyright infringement. Um, [11:14] it didn't want to, because it doesn't, yeah,  it doesn't. But Claude's models will do it [11:19] without thinking. So that's why I've been using  Claude for this use case. Yeah, interesting, [11:24] um, that's funny that Claude is actually more  open than GPT, it's the other way around. So, [11:29] it's task dependent. Um, the open models don't  care either, like GLM 5.2 will do it any time. [11:36] So let's talk about hooks a little bit. I actually  watched a couple other Codex videos, and I [11:40] actually like how they're more educational. Even  the hook is like, here's what I'm going to cover [11:45] in this video, as opposed to, you know, Codex  is like the most freaking amazing, like super [11:50] hype stuff. So is that like a choice, or, yeah, I  kind of feel like you have more substance. Yeah, [11:56] I mean, I try, when I make videos I try to make  it about the topic rather than any specific tool, [12:04] right. I always like to make it clear that even  if someone uses Claude Code, they're going to get [12:09] value out of the video. And so the best video  that I did on Codex, which got 300,000 views, [12:14] which I think is the one you're referring to,  um, that was like the nine ways to use Codex. [12:20] But I made it clear in the beginning, by the way,  you can do all of this in Claude Desktop as well, [12:25] most of the things you can do in Claude Cowork as  well. And I like to focus on the specific outcomes [12:31] that they'll get from the video, um, which I  think is really useful to establish in an intro. [12:36] So okay, so if you have some tips for, you  know, for like newbie YouTubers like me, [12:40] talk about the outcomes first and then get  into the content, right, is that kind of, [12:44] yeah, talk about the outcome. In fact, I would,  I don't know, did we film the intro already, [12:48] or do you do a separate intro? Are you going to  film a separate intro? I usually cut a separate [12:54] intro from the best parts of our conversation.  Yeah, okay, that's what I do. Yeah, yeah, okay, [12:59] that makes sense. And then do you do like  another little intro where you say, like, [13:03] today you're going to learn this, this, and  this, or do you just kind of go straight into [13:06] the video? When I was introducing you,  like, that was kind of my intro, like, [13:11] talk about how you're going to use it  to run your creator business and so on. [13:14] Yeah, I mean, for me, if I were you, I would be  doing testing. Like obviously you want to get [13:19] really good guests that are really knowledgeable  on certain things, and that's something that you [13:24] can, and probably already are, trying to improve,  but I would just test different intro types until [13:29] you find the winning format. You will find that,  um, you know, there's actually a really good [13:40] acronym, it's called BRENS, which is big,  relatable, easy, let me see, zoom in here, easy, [13:53] uh, new, and safe. The more of these you can  include in your intro, the better, right. Big, [13:58] like, there's a brand new model that just came  out, it's the best in the world, make it big. Um, [14:03] relatable is you find some theme, so,  like you say, like there's this new model, [14:07] it's absolutely insane, it just came  out and it's going to affect everyone, [14:10] or it's going to go into every business, right,  that's relatable. Uh, easy is, like, by the way, [14:16] I have a six step method that you can follow that  will guarantee that you will be successful, right, [14:22] like when you introduce some sort of methodology,  here are these nine tips that you can follow that [14:28] will guarantee, easy. Uh, new, obviously people  like things that are new, so like this brand new [14:34] model is also big and new. And then safe is the  one that's really important. Um, safe, is, good, [14:42] is important, this is only really important for  videos like over 30 minutes, it's like you want [14:47] to make sure the user feels safe watching, that  they're not going to waste their time watching [14:51] the video. And the way that you do that is you  say, like, "By the way, this guy's an expert, [14:55] he's going to talk about these things, and I  promise you by the end of the video you will [15:02] get value out of it." And you want to make sure  that they feel like their time is, yes, got it. [15:09] And how do you think about, so do you start with  the hook, or how do you think about the title and [15:13] the thumbnail? Right, because some, I don't think  about any of that, I think about that after. So, [15:16] like, I think about the thing that I want to talk  about in the video, which is like Codex tips, [15:21] right. This is something that I would want, like  I don't do a lot of research on what I should talk [15:26] about, I feel it in my heart, like my gut tells  me what I should talk about, and for 90% of the [15:32] video it's just me, like I'll outline it, but  it's just me getting into the weeds and talking [15:37] about how I actually use this tool on a day-to-day  basis. But then the last 10% is like, okay, how do [15:43] we get people to watch it, and that's where the  science goes in. And so, um, I start with a list [15:49] of everything that I'm going to talk about in the  video, and I just outline that, and then I film [15:53] the whole video, and then oftentimes I'll film the  hook after. Um, like a completely separate video, [16:00] because sometimes I'll go off on a tangent in one  of my videos, like some of my best videos I end [16:04] up on like a 30-minute tangent, [laughter], but I  want to make sure that my intro talks about that [16:10] thing that I do. So that's why I just do my intros  at the end, and I use AI to analyze everything I [16:16] talk about, and then that informs my transcript,  and then I'll have the YouTube researcher say, [16:21] like, "Here's the transcript for my video, I need  a good intro, scrape YouTube until you find a [16:27] good format, and then copy it and make it about  what I'm talking about," if that makes sense. [16:33] Oh, that's super interesting, dude, 'cause like  I know a bunch of other YouTubers who make the [16:38] thumbnail first, and if the thumbnail is not  good, then they would not even make the video. [16:41] But I guess, um, yeah, well, I mean, maybe, you  know, like for me the most sustainable strategy [16:47] for content is just to keep doing things that I  like. And if I let packaging and how many views [16:53] it'll get dictate what content I make, then it  just becomes a soulless game to me. Like I would [16:59] much rather make what I want and then package  it after the fact, than to just be, you know, [17:06] only about viral ideas. Then I would end up, you  know, it's like, why not, then I'll end up making [17:12] dance videos on TikTok, which is something I  don't [laughter] want to do, you know. Yeah, [17:16] yeah, so yeah, yeah, yeah, you got to think about  what gives you energy, right, and follow your own [17:19] curiosity. If you can make content following your  own curiosity, that's the best, you know? 100%. [17:24] Yeah, yeah, and then, um, to kind of continue  talking about Codex, you know, like I have a [17:30] lot of skills in here, um, one thing that's really  fun to do is I use Excalidraw in my videos a lot. [17:36] A lot of people watch my videos, like, I diagram  things out a lot of the time. So there's a skill [17:42] called Excalidraw diagrams that you can use,  and you can literally just tell Codex, "Hey, [17:46] I need an Excalidraw diagram skill," or "make  an Excalidraw diagram," and it'll just make it, [17:51] and then the link, you can just paste it into  Excalidraw. Um, it'll actually open it in the [17:57] browser, and it makes really good diagrams of  different ideas. That's one that's really useful. [18:03] I also have a YouTube thumbnail skill. This  one will get me into trouble a little bit, [18:08] because literally my YouTube ideation strategy,  when I have to come up with a YouTube thumbnail, [18:14] is I will just scrape a hundred thumbnails that  performed really well and put my face on it, [18:19] and then I'll ask AI to just, like, replace  those, with my face, and then I'll have AI edit [18:26] the color. Um, which, some people would get mad  at me for that, but that's been a really useful [18:34] skill. Um, and yeah, I just have a ton of things  that I've ended up using over and over again. [18:41] Um, also there's a Typefully skill. So if I get  an idea for content, um, this can control my [18:49] Typefully, which is connected to my Twitter.  So it can, like, really quickly draft up my [18:53] tweet ideas. Like if I'm working on something,  or come across something, I'll be like, "Hey, [18:57] draft a tweet on this." And then every Monday I  check my Typefully and fire off those tweets. Oh, [19:03] interesting. And so what I'm trying to get at  is, like, all of these skills did not come from [19:10] me sitting down and thinking, "Okay, what skills  do I need to create?" I just use AI to do stuff, [19:16] and then everything that's useful that I  do over and over again, I just tell Codex, [19:20] "Turn this into a skill," and then over time  I get in the habit of using certain skills, [19:25] and sometimes I'll stop using skills. But that's  kind of the thing, and the reason why I love Codex [19:30] so much is they have so many built-in plugins,  and their in-app browser is really cool. Um, [19:35] in fact, right before we were filming this  video, I was working on a short form script, [19:42] and the way that I write all my short form  scripts is I just start a new Notion page, [19:47] um, in Codex. And so I have this video database,  this is where I keep track of everything, right, [19:57] this is my long form videos, here are my short  form videos that I keep track of. Because it's [20:02] in-app browser is becoming a full in-app  browser, I can use it directly inside Codex. [20:06] And so what I want to show you really  quickly, right, I could say, "Hey, I [20:13] want to make a short form video  on the latest, uh, Meta Ray-Bans. [20:23] Please add a new short form in progress to Notion  and give me the link." Right, I'll often do this, [20:32] I'll usually say more stuff, like more ideas. I'm  just trying to illustrate the point that, like, [20:36] right now the agent is going to go into my Notion,  it's going to add it into the database correctly, [20:41] and then it's going to give me a link, and I'll be  able to open that up directly in the app, I won't [20:46] have to leave Codex, it'll just open directly  inside Codex, which is such a fun workflow. Like, [20:52] I love not having to leave the app, I get lost  when I switch from tab to tab. You see here it [20:57] gave me a Notion link, I can just click open in  browser, and look at that, it added this. So it [21:03] added the notes in here, which is fine, but it did  create the Notion database. And I can say, like, [21:09] um, "Please make the outline in the body,  and please make it with a good hook." I would [21:22] normally use, um, say that this will be a huge  opportunity if they take off, for creating [21:38] apps on these glasses, I don't know. And yeah,  you just control the Notion, which is really fun. [21:45] And this is how I use Google Docs and Notion,  and my calendar as well, like I'll just say, [21:52] give me a link to the calendar event, and I'll  open it in the in-app browser so I can just see [21:55] it right there. And it can also write all of  my emails, which I have it do all the time. [22:01] Yeah, I have it write my emails too. Yeah, okay,  so so basically then, one day you can just wake up [22:05] and, like, you have a bunch of scripts already,  and you can just batch process, you can just [22:10] record them in batches, right, or something. Yeah,  yeah, I think that's where everyone's brain goes [22:15] at first, is to like batch and create more. What I  want is better, I like whenever I create content, [22:22] I'm always thinking about how to create  stuff that's better. And what I want is, [22:26] like, I want multiple tabs to open up, you know,  in the background, of relevant things that might [22:31] be useful. Like Codex would be like, "Hey, by  the way, I opened up this link, you may want to, [22:36] in this one specific line, you should do this."  And so I will actually use it to just get AI to [22:42] fact-check everything that I'm talking about in  this video, or I'll just send it off and say, [22:46] like, "Hey, can you go and check this creator's  videos, he has really good pacing in his videos, [22:51] can you please figure out a way that we can  implement that into this video?" It'll go [22:55] off for 10 minutes and come back. And that's  kind of how I see this, is, like, you can do [23:00] research so much faster, so it allows you to put  more total research into one video. Um, because [23:06] I think the moat is quality over a long period  of time versus batching. Batching will make it [23:14] soulless over time. So that's kind of how I think  about it, and I would encourage others to do so. [23:18] Got it. At this point in your creative journey,  like, how many short forms are you publishing [23:21] every week, one or two that are high quality?  Yeah, I mean, I would like to create a video [23:26] once a day, honestly, like five days a week I  would love to create a high quality short form [23:30] video. Um, yeah, as of late I'm trying to get  my long form to like four videos a week first, [23:36] and I think four long form, five  short form is my goal for sure, [23:41] I think that would be good. And it  won't be too hard. I make one or two [23:46] videos a week and I'm already exhausted, so  [laughter], I have to use Codex more. Yeah, [23:51] I use Codex more. Yeah, I don't know, I like  it, it's fun to make content, I don't know. [23:57] Yeah, dude, I have to ask you, okay, so let me  ask a few more questions about the skills that [24:00] you kind of showed. So the Excalidraw skill,  do you use that after you have the script, [24:05] or before you have the script, like,  you make the diagrams? Yeah, I mean, [24:09] I think I will just blabber, so, 99%, I'm not, I  know I'm typing now because this is a podcast, and [24:15] I don't want to use Wispr Flow, I use Wispr Flow  for everything. And so what I will do sometimes, [24:21] um, I got these AirPods Max because they have  a better mic, I'll just turn on Wispr Flow [24:27] and I'll just walk around for 10 minutes,  just speaking all of my ideas, like, "Oh, [24:32] I'm gonna make a video on Codex, the things that  I've been thinking about this week are this, [24:36] this, and this, and this. I want you to  create an Excalidraw diagram of this, this, [24:40] and this." I have, you know, um, and then  I'll literally tell it, I want nine slides, [24:46] and my Excalidraw diagram skill does it in a very  specific style, which has just sections to them. [24:53] And yeah, I'll just blabber for 10 minutes,  and then I'll be done, and I'll just run it, [24:58] and then I'll come back, and it will have like  80% of the diagrams that I'll actually use in [25:02] my videos, and I'll edit them. I'll take 20, 30  minutes to edit them, but it's like a really good [25:07] starting point. Um, it just organizes my ideas,  and yeah, huge fan of Wispr Flow for that reason. [25:14] Okay, so so okay, so the slides actually  kind of frame the rest of the video, [25:17] right, because you want to cover like eight  slides, and do some demos and stuff. Yes, [25:21] yes, got it. So then I have to ask you about  the thumbnail skill, because I have a lot of [25:24] trouble using AI to make thumbnails for  some reason. So, sure. So what I found is [25:28] that it's actually really bad at changing your  face, or maybe I don't know how to prompt it, [25:32] if I give it like a picture of me smiling and  tell it to change to a frown or like a surprise, [25:36] it totally distorts my face, dude, so, like,  do you mostly apply your face afterwards? [25:41] So, um, one of my most popular tools that I talk  about that I use with Codex is a tool called [25:48] Paper. Paper is an AI-native Figma, and when I say  AI-native I mean that it is an application built [25:56] for AI first, this, they basically took a lot of  the good stuff about Figma and made it really easy [26:03] to connect to any tool. And so when I'm using my  thumbnail skill, I will say, "Hey, I need you to [26:11] scrape," this is a short form video that I'm  releasing in a couple days, uh, where I ask AI [26:17] to scrape Alex Hormozi and Dan Martell's YouTube  thumbnails and place them on the Paper board. [26:22] Codex will go off for 10 minutes finding his most  popular videos for each one, and then it will [26:28] place them on this board. Now all I have to do is  I can come here, and since I have images of me, [26:32] right, this is a thumbnail of me, I can duplicate  it, and then I will just come here and do a lot [26:37] of it manually. So, like, I'll grab this one of  Alex Hormozi, I'll drag this over here, and then, [26:43] um, Paper has AI image generation built in.  And so, watch this, I can just go like this, [26:49] select this, uh, "Please replace the guy with  a beard with the guy that doesn't have a beard, [26:56] just replace the guy, except make his  shirt white, and make him wearing a chain." [27:04] And then, you want to wear a chain? Yeah, just  because, you know, this podcast is important, [27:11] it's important that I'm wearing a chain in  this thumbnail. Yes, it's quite important. [27:15] But it looks like it's generating multiple  thumbnails. Oh yeah, I just spam-pressed enter, [27:20] so it's creating uh, seven versions of it. And  then I'll just kind of select the right one. [27:25] And you can see kind of my natural progression,  right, you can see that this says "game changer," [27:30] I took this and then I converted it into  one for me, um, and here are my thumbnails. [27:35] And it's kind of fun to see the progression  of the thumbnails that I've created, over the, [27:44] you know, I created this one right here, which  was based on this one. This is Andrei Jikh, [27:48] he has great thumbnails. And so that's kind of  how my thumbnail process, I don't like to think [27:54] about thumbnails for that long, I despise making  thumbnails manually, so this is great for me. [27:59] Oh dude, this is amazing, so okay, so it basically  just replaced it, it kept his hand, it looks like, [28:05] yeah, it kept his hand, I could have specified  to change it. But then at any time, you know, [28:10] like what I like to do, I'll zoom in on this one,  right, you know, this is not bad. I can come in, [28:15] oops, and come in again and make it small, I  like to make it smaller. And then I'll just [28:20] make more variations, like, "Hey, please smooth  out the face without changing what the character [28:26] looks like." And then also add, like, an outline  glow that is kind of whitish orange around him to [28:33] separate him from the background, and make the  green line brighter. And then I'll just kind of [28:41] iterate on it. Um, and this gives you this kind  of starting point. But the useful part about [28:49] Codex is Codex has full control over Paper. Um,  and so it can just add useful thumbnails to it, [28:55] because again, you know, going back to what you  were saying about, uh, what did you call it, [28:59] chain prompting or something. Um, yeah, it can go  off, do a thing, and then it can use a different [29:04] skill and place it on the board, because  it has that access. And yeah, and do you [29:10] have some sort of, uh, okay, so right now you're  just prompting in Paper, right, with Wispr Flow, [29:15] yeah. Do you have some sort of like a style  guide, or, you know, your text is always white [29:18] or something? You don't need that. Like, so, I  actually have a YouTube thumbnail designer too, [29:25] and he uses AI all the time, and he helps me  with this, and I'll actually send him these. [29:30] I could use these and it would do pretty well.  I love having someone who is constantly doing [29:35] the A/B testing and everything. Like, I do  a lot of A/B testing because, like, why not, [29:41] it's worth paying for, because, yeah, like if  your video does twice as good, that's a big deal. [29:47] See, it kind of made the outline here. Um, but in  terms of style guide, if I want it to look like [29:53] something, I'll just use it as a reference. So  here I have, somewhere I have all my thumbnails, [30:00] um, I don't know where I put it, but, like, let's  say I wanted it to look like this, I would just [30:04] copy this thumbnail, bring it up here, and then I  would say, and you can very easily copy multiple, [30:11] so I can literally copy, select this  one, and then I can select this one, [30:15] and I could say, like, "Please make the guy  look more like this, and make the text match [30:20] the one that says Miss Mythos and Fable," and  that's how you do style references, you know, [30:27] and you don't need to do any prompting, you  just, um, you know a lot of these image models [30:33] are called Nano Banana. And so you can just  think of prompting, you can prompt in visuals, [30:38] you can say, "Look, make it look like this,  make the text look like this." And so that's, [30:41] that's why I like the board, it's like I can  just be like, "Oh yeah, I want it to look like [30:45] this." And then I'll just use AI to make it  look more like that, if that makes sense. [30:49] Dude, this is like super educational for me, okay,  I think I'm going to become a Paper customer now [30:53] too. Yeah, you should 100% do this. Yeah, you  should 100% do this, and get all of your top 10 [30:58] performing thumbnails and toss them on the board,  right, you don't need to use Codex, if you had [31:01] my YouTube thumbnail skill, you could just say  Codex, do that, and it would be done. All right, [31:06] dude, well, you definitely blew my mind with  this thumbnail thing, this helps a lot. Yeah, [31:11] I don't even know what other question to ask  you, that's something, so mind-blowing, you know? [31:15] Yeah, I mean, and then, I think what it all comes  down to, if I were to summarize, is just find [31:21] useful workflows and turn them into skills, and  the more skills you use, the more you'll realize [31:28] that you can mix and match. Like this here is a  mix between my YouTube researcher skill, I just [31:32] had an idea one day, I'm like, "Hey, can you pull  all these images?" And then it gave it to me in [31:37] this list, and then I had to go into my downloads  and see it, and I couldn't visualize it. I was [31:41] like, "Oh wait, you can add it to Paper." And,  because I'd been using the Paper skill to create [31:45] designs, right, this is, Paper's not made for  images really, it's made to, like, you can ask [31:50] it to create full designs, and it'll do that.  And then I was like, "Wait, can you add this [31:55] to Paper?" And then it unlocked this whole new  workflow for me. And you'll, the biggest unlocks [32:01] you'll find when using AI will come from this  experimentation process, that's just really fun. [32:08] Do you look at the skill files manually,  no, you just look at the output, right? No, [32:11] I've never looked at a skill file once, like, it's  test-based, it's like, ask AI to do a thing, use a [32:18] skill, if it does properly, great. If it doesn't,  you tell the AI, "Hey, you didn't do a good job," [32:22] and then, uh, "please change the skill so that you  don't do that again." Then you want to go to a new [32:28] chat, so the context is cleared, so you can test  it again. Test the agent in a new chat, see if it [32:33] does it successfully. If it does it successfully,  great, right, and once it does it incorrectly, you [32:37] update the skill, right. That is the way better  way to do it than to manually edit the skill. And [32:45] there's a lot of people saying, like, "Oh, you  need to write your skill like this, or else it's [32:49] not going to perform well." It's like, first of  all, that's only true with that model, and it's [32:53] only true with that model today, right, as soon as  there's model updates. So I think you just want to [32:59] be outcome-based, right, hold your AI to a certain  standard, and then you can turn it into a skill [33:05] and improve it by measuring their outcomes. That's  basically how I do it for most creative workflows. [33:10] Yeah, and I feel like a lot of these skills  are just super personalized for you, or for me, [33:14] that's kind of the power, right? I feel like  that's why skills is kind of taking over a lot [33:18] of the SaaS stuff, because, you know, you can't  make a SaaS product as personalized as this. Yes, [33:24] and actually, I will caveat that with, this  is for my creative workflows, right, they're [33:28] very low risk. You know, if we were doing some  mission-critical thing, which we actually do, like [33:36] when we're dealing with clients, or when it has to  do with anything that has to do with payments or [33:40] sensitive documents, yes, we'll go in and analyze  the skills, and there's kind of like a process, [33:44] we'll get someone else to analyze the skill  if it's super high impact. But for my creative [33:49] workflows, just realize that there's no risk, just  let your AI cook and tell it to do a better job. [33:55] Um, and then, yeah, the next thing for  me is figuring out how to make this [34:02] for my team. And that's what I think  Codex and Claude, or Anthropic and OpenAI, [34:07] they're trying to figure out, it's like, how do  you enable this process with teams? And that's [34:14] going to be a really interesting situation. And it  feels like Anthropic's going after Slack. Um, and [34:20] so that's kind of the next wave, maybe, I'm not  fully sold on Slack as the interface, but there's [34:28] something really interesting going on there that,  um, yeah, I find it to be really interesting. [34:33] Yeah, I think you're right, I think, like, as  good as the Codex app is, it's like a single [34:36] player experience, right, other than, like,  putting PRs up and stuff, it's pretty much a [34:41] single player experience. It is single player,  and it's really hard to create these creative [34:45] workflows in a multiplayer way, which is something  I think about a lot. So yeah. Do you have, like, [34:50] a bunch of automation setups that kind of do work  for you? Like, yes, I have a lot of automation set [34:55] up. I don't use Codex for my automations, simply  because I don't want my computer to have to be on [35:01] to run those. Um, and so I have Claude agents.  Yeah, and so, like, I have a lot of Claude agents. [35:06] In fact, I created a Peter Yang bot right before  this episode. I didn't, I was going to show you, [35:12] but, like, I created it in like five minutes.  So, like, I have my chief marketing officer [35:19] agent that I have in Slack, he sends me, um,  this guy, he sends me updates. And then I have [35:28] another guy named Content Man, who analyzes a  bunch of other people's content, and it'll just [35:33] give me video ideas for short form content. And  so I have this little channel called My Agents, [35:38] and I have three agents in here, [laughter], one  of them is Peter Yang bot, which is trained on [35:42] all your YouTube videos, but I didn't finish it  in time to make it good. No problem, no problem. [35:48] So so these agents are based on your own  product, Chorus, or is it based on, like, [35:52] uh, something else? No, we've been testing it with  a bunch of different products. These three agents [35:58] are, um, created on Chorus, um, but there's a lot  of products out there that you can create agents, [36:03] and I just recommend trying it, because  no one's figured this out yet, you know, [36:07] no one's figured out how to create a team of  agents running in Slack that your whole team can [36:16] use. It's like a really tough problem to solve.  The way Anthropic is trying to do it is they're [36:21] trying to build like a little mini, for lack of  a better term, god that knows everything, where, [36:27] uh, other companies, including us, we think of  it like each one will be team members. So you'll [36:33] tag your marketing agent, and you'll also tag  your developer agent, company developer agent, [36:37] um, and you'll kind of have a relationship  with the agent, um, like each person, it'll [36:44] react differently to each person based on what it  knows about the person. There's like a whole bunch [36:49] of problems that need to get solved. But you  can very easily export your Codex skills into, [36:56] um, these Claude agents, which are really fun. And  so yeah, it's just been a lot of experimentation. [37:01] But Claude agents, uh, tell me, if I want to tweak  something where I'm signed into some sort of app, [37:06] Claude agents can't do that, right, or, like,  they need your credentials. Yeah, I mean, there's, [37:10] with browser use, I predict by the end of the  year you'll be able to be signed in to all of [37:17] your apps, and your agent will be able to use a  browser, and you won't even be aware of it, like, [37:21] it'll just be able to use a browser, um, almost  perfectly, and it'll be able to be signed in. Um, [37:28] and I think this is something that the  big labs companies want to do as well, [37:30] like a browser being able to control a browser in  the cloud. Um, is much better for automations and [37:36] cron jobs as well. Like, I don't know if  you have any automations set up in Codex, [37:40] but it is kind of annoying. Um, like I have  none set up right now. I have Codex running [37:44] on my little Mac mini, I have some automations  there, that's kind of how I get around it. But, [37:49] um, yeah. Yeah, that makes sense. But then, if  it's a separate computer, it doesn't have access [37:53] to your core files, which might be important.  Yeah, like, it's a pain in the ass. Yeah, [37:59] pain in the ass in so many different ways. Like  every agent type has its trade-offs, um, which [38:07] we're seeing right now, we're trying to, all the  companies are trying to figure out the right way [38:11] to do agents, and, yeah, maybe it'll take another  six months to figure out the best way to do it. [38:16] Cool, all right, dude, well, let me ask you one  last question on kind of the overall business, [38:21] right. So I feel like the content,  from one content creator to another, [38:25] I feel like just growing the audience and getting  more sponsorships, that's kind of a surefire path, [38:30] because there's a lot of AI companies who  want distribution and attention. Sure. And [38:34] then there's also a lot of demand, you know, like  you said, there's a lot of demand from companies [38:37] that want to build agents, train their employees,  right, [laughter], but, like, in terms of actually [38:42] building a product, that's a lot more difficult,  just being like a SaaS product. It is a lot more [38:46] difficult to do. And I do brand sponsorships, and  I run mid-roll ads, and it's a great business, [38:51] I mean, it's an insane business. Um, yeah, and  it is, I could very easily just kind of ride off [38:59] into the sunset and make a very good living making  content. Um, I think my content would run dry if I [39:05] did so, because I learn being in the trenches, you  know, like, I enjoy, the only reason I know a ton [39:13] about Codex is I spend time with developers  every day who teach me really cool things, [39:17] you know, like every day I'm learning three  or four things, and we're actually, you know, [39:22] in the content that I make, I actually want to  help people implement agents into their business. [39:27] You know, I want to be a very useful creator and  a motivating creator, and I don't think I could [39:32] do that if I was just making content. Um, and so,  yeah, I think doing, like, especially consulting [39:39] stuff, I think it really does help, especially  if you build a really talented team. You know, [39:44] my goal is to only be making content, I can still  contribute to the business through top of funnel. [39:49] Um, but ideally, all I'm doing is making content,  and you just have to surround yourself with a [39:55] really good team, and there's a lot of talented  people who would love to partner with you and [40:00] do all of the work in creating a product,  and I think you would learn a lot from it, [40:04] um, or anyone in your position. Um, there's  just a lot of upfront work that goes along. [40:10] Yeah, I don't want to just, same here,  I don't want to only make content or be [40:14] like some sort of news aggregator, you know, I  want to actually learn how to build the stuff, [40:18] talk about the building process. So, 100%, I mean,  yeah, like, you know, Peter Levels on Twitter, [40:24] yeah, you know, like he could never be a content  creator, like, he's only interesting because he [40:29] talks about the things he's built and the  experiences he goes through. And so that's [40:32] kind of how I think about it, it's like,  you know, that's the creator I want to be, [40:35] is someone actually trying to build and use this  technology. And yeah. Yeah, he's one of the few [40:40] folks who actually have a bunch of successful  SaaS companies, because his distribution is so [40:44] high, [laughter], and a lot of failed companies  as well, he's failed like 40 times. Um, [40:52] he just takes a ton of swings at bat, which is  also something that makes him really interesting. [40:58] Cool, all righty, well, I think people know where  to find you online, do you want to do a pitch [41:02] for your new product, Chorus, or should we talk  about it next time? I mean, yeah, we're building, [41:07] you know, Claude Tag just came out, Claude, uh, is  a product that allows you to tag Claude in Slack. [41:13] Um, and then you have all your skills built into  Claude. We're building a product that allows you [41:17] to use any model. Um, you can message it  in iMessage, you can message it in Slack, [41:21] you can add it to group messages, which is what  our users like the most. And we're constantly [41:25] iterating, trying to help businesses. Right now  I'm not even really selling the product as much [41:29] as selling our services, because companies want  to work with us, and then we're using Chorus [41:34] to implement agents for their business, and  that's been just like an insane driver for us. [41:38] Cool, all right, well, you're an inspiration, man,  thanks for being so open and sharing your whole [41:42] process, and I personally learned a ton, so thanks  so much. Amazing, yeah, thanks for having me.