信源说明：本文基于 Matt Wolfe 频道视频《He Put Hormozi, Bezos & Buffett Inside Claude》（视频 ID AtjFk9f44HE，发布于 2026-10-29 北京时间凌晨，全长约 38 分钟，转写共 20 段、每段 120 秒，覆盖 0–2289 秒）英文转写整理，为「Show Me Your Workflow」系列第二期（第一期嘉宾转写疑作 Bolavos Sudo，展示了其 God’s Eye View）。文中全部做法、数字与引语均为 Matt Wolfe 与嘉宾 Nate Herk（转写疑作「Nate Hirk」，按其公开身份校正）的自述，未独立核实；归属按「Nate Herk 称／Matt Wolfe 补充／可以推断」分层；ASR 存疑词以「转写疑作」标注。
先说结论
这期节目没有一个新模型、没有一条大新闻，但它可能是近期最值得普通人照抄的一期：一个独立 builder 用 Claude Code 把「内容剪辑、决策顾问、社媒运营」三件事各跑成了一条日产级流水线。真正的干货不是工具清单，而是一条反复出现的心法——给 AI 的最好起点不是空白提示词，而是「已验证的成功样本」：Nate Herk 让模型逐帧分析 27 条人类剪辑的爆款 reels 建成模式库，从此 AI 剪出的片子与他剪辑团队的风格一致；他把 Alex Hormozi 的 123 个视频、超过 150 万字言论喂成「大脑」，让子代理议会真的像那个人一样提问。三个工作流的共同结构是：人退到导演位，把审美与判断沉淀成可复用的技能文件，其余交给代理。对读者的直接信号：内容行业的「一人公司」门槛又被踩低了一截，而新的稀缺物是「什么是被证明有效的」这个金标准本身。
心法：把「金标准数据集」从评测搬进创作
Nate Herk 的出发点是一个对大模型的清醒判断（120–240 秒段）：模型本质上在预测下一个词，所以它默认给出的是「多数人会怎么想」的平均答案——用来创作时就是平庸。解法不是换更强的模型，而是给它非平均的参照物。他直接借用了生产环境部署 AI 的逻辑（240–360 秒段）：上线前跑 evals、拿金标准数据集（golden dataset）当靶子，那为什么创作不能反过来「先定金标准再开工」？于是剪辑技能的第一步不是写提示词，而是让 Claude Code 把 27 条由人类剪辑、已被市场验证的 reels 逐帧看完（360–480 秒段）——每隔几秒截图，拆解镜头切换频率、B-roll 用法、动效、甚至「我的语音、音乐、音效的音量差」，最后沉淀成一个模式库（pattern bank），再转成技能文件反复复用。他把这条原则上升到决策层面：做任何事之前，先找一个「已经做成过这件事的人」，拆解他怎么成的（240–360 秒段）。
工作流一：AI 剪辑流水线——从 27 条爆款到「AI 剪辑团队」
有了模式库，剪辑流水线才立得住。Nate Herk 描述的当前产出能力（120–240 秒段）：一个早晨产出约 20 条 reels，选题、脚本、B-roll 抓取、AI 生成视频全部自动，他只需对着提词器念。技术栈是 Claude Code 主控、后端渲染（转写为 Hyperframes，未确认）加浏览器自动化去抓 B-roll、截屏、录屏，AI 视频用转写疑作 Kling AI 或 Higgsfield 生成（360–480 秒段）；他甚至没打开过 Blender 的界面，就让 Claude Code 接管做出了 3D 场景——Matt Wolfe 呼应说自己在三个月里用 AI 做的 Blender 东西比一辈子都多。
两个工程细节值得抄。第一，视觉理解的替代方案：Claude Code 并不能「看」视频（600–720 秒段），Nate Herk 的做法来自他做网页设计的经验——要求代理对每个元素截图、滚动到底逐一截图检查越界——把「逐帧截图」烘焙进所有需要视觉判断的技能，YouTube 长视频干脆每帧都截。第二，音画同步的坑：用 ElevenLabs API 转写拿到词级时间戳（比如这个词从 0.09 秒到 0.17 秒），动效得以精确卡点；但按词切仍会削掉词尾，于是技能再叠加一层看波形、等「平了」再切（720–840 秒段）。长视频剪辑还内置了「观众视角」：检测鼠标静止多秒判断该切全屏幕人像、为可读性自动放大局部、发现讲解断档主动上报（480–600 秒段），并用 LLM-as-a-judge 逐步积累更好的裁判，让技能随反馈变好。
边界他也认账（240–360 秒段）：音乐与声音设计仍是最弱环节，「人对这件事还是会做得好得多」；他的应对是把喜欢的音效、音乐、热门音频存进自建的素材库（Herk Edit Kit，转写疑作）持续喂给技能。两人的另一个坦白：这套东西让人变「懒」了——几分钟的手工活现在不想干了；Matt 的圆场是，人的心智确实沿价值链上移，去做编排与更大方向的思考（480–600 秒段）。
工作流二：Roast Room——把专家顾问团工程化成子代理议会
第二个工作流解决「拿不准的决定找谁吵」。Roast Room 是一个跑在 Claude Code 里的代理议会：你抛一个带选项的决策问题，每个议员先亮开场立场，再互相辩论、投票，最后给出裁决和一份 PDF 概览（840–960 秒段）。议会成员是 Naval、Codie Sanchez、Leila Hormozi、Sara Blakely、Gary V、Paul Graham、贝索斯、巴菲特（Warren）等，每个成员是一个独立的子代理 markdown 文件，有各自的「大脑」（语料库）、「视角」（怎么想、怎么辩论、绝不做的事），并且每人只拥有一个专属问题——Alex Hormozi 管定价与成交（「人们会付钱吗？付多少？怎么成交？」），Naval 管杠杆（「这建的是你拥有的杠杆与资产，还是在出租你的时间？」）（960–1080 秒段）。
人格的保真度来自语料的工程化（1080–1200 秒段）：Hormozi 的大脑消化了 123 个视频、书籍框架、文章、X 与 LinkedIn 帖子，节目称超过 150 万词直接出自本人，再提炼成热座问题、三句话世界观、八条决策规则、书、框架、立场与故事。可视化的头像用 GPT image 2 生成，还有一个「new roaster」技能可把任意新人物研究后加进议会（节目里开玩笑说该把马斯克加进去当「搅局的」）。架构上有个务实的取舍：这套东西曾用 agent team 跑，但成本更高、价值增益很小，于是退回子代理方案（960–1080 秒段）——多代理编排不是越复杂越好。
现场演示挑了 Matt 自己的焦虑（1200–1320 秒段）：他的 AI 资讯 newsletter 有 26 万订阅、收入全靠赞助，而他担心未来「读 newsletter 的都是 AI 代理，没有人类眼球」。议会自动生成五个选项（维持现状／为人类重建／按效果重售赞助／为代理设计／出售代理触达并转移渠道），并自行调整了议员阵容——换下 Dan Martell 和转写存疑的 Gino，换上 Naval。辩论过程可视可查：Gary V 炮轰 Hormozi「读者没问题，坏的是按打开量定价的赞助报价」；Paul Graham 也反对 Hormozi，贝索斯与巴菲特互有分歧（1440–1560 秒段）。投票结果除 Warren 与 Alex 投 C 外全部投 B，最终裁决（1560–1680 秒段）：主推 B「为人类重建」，同时旁跑 C「按效果卖赞助」——理由是 B 是唯一检验「读者到底要不要 Matt 本人」的选项，而按打开量计价的赞助正把整个商业模式押在「代理不来读」的假设上；节目按假定数字估算约 7.8 万美元年收入处于风险中。Matt Wolfe 的点评值得记住：这套东西的产出（辩论记录、评分、行动清单）在没有可视界面时完全一样——界面只是「更好玩」。
工作流三：99% 自动化的 Instagram 账号
第三个工作流是一台内容自动售货机。一个运营约一个月的 Instagram 资讯账号，Nate Herk 称自动化程度 98–99%：每 4 小时自动找素材、生成 carousel、发布；他也可以随时把一条 X 帖子的链接丢给 Claude 说「make a carousel for this」，约 5 分钟后得到成品（视频素材更久），确认后自动上线（1680–1800 秒段、1920–2040 秒段）。生成的深度超出「排版」：代理会去读帖子下面的整个回复串、读相关文章、找到事件当事人的视频片段（比如找到 Dario Amodei 的发言片段连音频一起嵌入）、用 AI 生成封面图（「抱着一堆钱哭的 Dario」）、把人物抠图分层且避免遮脸（1800–1920 秒段）。 Engagement 侧用 Botato（Sabrina Ramonov 出品的社媒自动化工具，转写先误作 Bloker）做多账号管理与 DM 自动化：评论区留言领资料 → 私信里「先关注再发你」，一条爆款帖收到 559 条评论（1680–1800 秒段）。
这套系统里最值得抄的是核查层：发帖前会确认信息为真（X 上假消息多），正式发布前还会再刷新一遍源帖——因为已经发生过「carousel 做好了、源帖却被删除或编辑」的情况（1920–2040 秒段）。所有发布记录进 Google Sheet 防止选题撞车。成绩单很克制：30 天约 2 千粉丝，累计人工投入 4–5 小时（1680–1800 秒段、1800–1920 秒段）；Matt 评价「30 天做到这个数已经很好」，Nate Herk 也承认这活儿以前是一个全职岗位。下一步两人现场脑暴：把每周表现数据拉回 Google Sheet，让技能看着数据自我改进——一个自改进闭环的雏形（2160–2280 秒段）。
三条「机制 → 二阶影响 → 读者启发」链条
- 金标准数据集机制：AI 输出趋近其参照物的分布 → 用已验证成功样本（27 条爆款、150 万词语料）当参照物，产出直接继承「被市场证明过的风格」→ 二阶影响：数据资产从「训练数据」下沉到个人技能文件，谁手里有验证过的样本谁就拿到质量下限 → 读者启发：动手让 AI 做事之前，先花力气定义「什么是好的」，这一步无法外包（Nate Herk 称之为整个转变里最重要的心态，240–360 秒段）。
- 技能复利机制：每次人类反馈（删掉烂帖并说明原因、LLM-as-a-judge 打分、波形切点修正）都被写回技能 → 边际成本递减、质量单调上升 → 二阶影响：工作流之间的能力可迁移（截图循环从网页设计搬进剪辑），个体产能从「天」压缩到「分钟」（一早 20 条 reels）→ 读者启发：值得把「我会反复做的事」全部沉淀成带反馈回路的技能，而不是每次重新提示。
- 自动发布的信任机制：全自动产帖 + 全自动发布会把错误也自动化 → 系统必须在发布前加「事实核查 + 源时效复核」两道闸 → 二阶影响：自动化程度不再是瓶颈，「核查与审美」成为人工保留地（音乐音效、最终删除权仍在人手里）→ 读者启发：设计任何自动对外输出的系统时，先设计它的刹车，再设计它的油门。
边界与反方
- 两个都没解决：Roast Room 没有语音（只有「摇头晃脑的嘴部动画」），音乐与声音设计仍明显是人类强项。
- 成本敏感：多代理编排曾因「成本更高、增益甚微」被降级回子代理；Berman 们在别的视频里反复强调的「每任务成本」在这里同样适用。
- 样本偏差：27 条爆款学到的可能是「这个频道受众的口味」而非普遍规律；Nate Herk 也承认 2 千粉「不算惊艳，但更接近真实」。
- 议会的群体智慧边界未验证：议员立场来自语料统计，「投票」更像多视角清单而非真博弈；节目未做任何对照评估。
- 所有数字（20 条/早、150 万词、559 评论、7.8 万美元、4–5 小时）均为节目自述口径。
归属与转写说明
本文全部事实与数字出自本集转写（含 26 万订阅、约 7.8 万美元年风险收入等 Matt Wolfe 自述口径，未独立核实）；Roast Room 对 newsletter 的裁决是演示性输出，不构成经营建议。转写校正：Nate Hirk→Nate Herk、Herc Edit Kit→Herk Edit Kit（疑作）、Alex Hermosie/Hormosy/Hormozy/Hermos→Hormozi、Layla Hermos→Leila Hormozi、Nival→Naval、Cody Sanchez→Codie Sanchez、Sarah Blakely→Sara Blakely、Carpathy→Karpathy、Mini Chat→ManyChat、a money ManyChat→a ManyChat、Bloker→Botato、Sabrina Ramenov→Sabrina Ramonov、Love and Labs→ElevenLabs；按原文保留并标注疑作：Key AI（疑为 Kling AI）、Bolavos Sudo、Hyperframes、Gino。附录转写已按上述确认项机械替换。
转写原文（修正版）
[0s - 120s] This channel is all about finding you the coolest AI tech tools, models, and perhaps my favorite thing, which is workflows. This video is the second in a new series that I like to call “Show Me Your Workflow,” where I bring guests and friends onto the channel, and then I mercilessly squeeze every last drop of value that I can from them. And by squeeze, I mean I get on a Zoom call and politely ask them questions about the cool stuff they built. And today, I’m talking with Nate Herk, one of my all-time favorite builders of cool stuff. He’s going to walk you through three of his favorite AI builds that he uses every single day: his automated AI editing workflow, his AI expert panel called the Roast Room, and his AI social media manager. And the best thing is, by the end of this video, you’ll be able to do all of this stuff on your own too. So, Nate Herk, show me your workflow. Let’s optimize that workflow. So, sounds like the first one we’re going to get into is like the AI editing workflow, which is this is something I’ve been hearing a lot about from a lot of people, and I’m really really excited to hear your take on this because I’ve tried to do some just like let the AI edit for me, and I don’t know, I always find that like I’m spending more time when I let it do it for me because I’m going back and sort of fixing a lot of stuff. So, I’m really really excited to hear your your editing workflow here. Yeah, this has been something that I’ve been very very surprised by in a good way because I think when all this started, I agree with what you just said. It was just like it was frustrating because it would take more time because the rendering would take a while, and then you would make a change, and then it would have to re-render everything. But once I really started to change the way I thought about how I was doing these editing workflows, as well as of course saving all of the knowledge and the feedback and building skill.
[120s - 240s] Around it, then this has truly made me so much more productive. Like it’s it’s it’s pretty nuts. Like just this morning, for example, I had worked this in combination with some other skills that I’ll talk about in a bit, but was able to just bang out like twenty reels. It did all the ideation, all the scripting, and then I just threw it on my teleprompter, read off. The cool thing is like it grabs all the B roll and it makes all the AI generated videos that go into it. So it’s it’s really a cool pipeline. But I wanted to start off here before maybe we look at like some some examples and stuff about how it was built. And this is kind of a wordy doc, but I just wanted to talk about the the way I think about not only this workflow, but in general the way that I build skills and workflows that I think is really a game changer. So it’s kind of timely with the whole you know reverse engineer anything repo that dropped and just went insanely viral. Oh yeah, the whole idea is that I’ve been trying to build skills around editing short form content for a while, and it just was never really there. It just always felt a little cheap. And AI on its own, it’s not very great at thinking outside the box. I mean, if you really think back to how it was actually like what it does under the hood, like it just predicts the next word, right? Like at the most fundamental highest level, which means that it’s going off of like what most people would assume or like what makes sense. So what I started to do was, you know, I had all of these reels that were proven. Now I can take this raw input where I’m speaking, right? Like I’m just able to record, even if I have mistakes in there, I’m just speaking. And then what I have right here side by side is the manual versus the skill now. So this is like a human editor had built this, and you can see this the styling is very consistent, the font, the way the motion graphics come in, the shimmers, the sound effects, and the music are pretty similar too. And that is one of the areas where AI I think still has a lot of work to do. Music and sound design, I think music and sound design is super super underrated in the experience for a viewer. And right now, like a human is just going to do that a lot better, but it is getting there. And I’m starting to get to this point now where my Herk Edit Kit is what I call it. It’s saved a bunch of the sound effects I really like and sounds that I like, like music, trending audios, and I’m just constantly.
[240s - 360s] Feeding in like when I would want it to use this one, you know what I mean. So it’s definitely getting better. But the point I was trying to make here is you can kind of see there’s like three main scenes, right? There’s me on the bottom with B-roll up top. There is B-roll full screen, and then there’s me full screen. And basically the way I was able to get to somewhere where I really liked this is I had it go do what was proven. And so before I continue scrolling down here, I think that’s just a really really important mindset to have when you’re working with AI is when you’re trying to have it help you with something, the best inspiration is is proven work. So like if you think about when you’re deploying AI in production, what do you do? Is you run evals and your evals test against the golden dataset. So essentially like why would you not start with that kind of golden dataset in mind when you’re building? Because then you know what you’re building towards. So whether it be an editing flow or whether it be even just like a business decision, if you go look at an example of someone who’s done exactly what you want to do successfully and has proven it, then break down how it works, how they got there, and I think that just if you zoom out and think about life in that way, like it’s changed the way that I I make decisions. But so real quick, back up on the one that you were you were showing there, you’ve got the two videos side by side, and it looked like it was finding pretty much the same B-roll, right? Like it’s the same video of Karpathy there. So the human one found the same B-roll as the skill, or did you tell the skill which B-roll to use? Well, in this case, what happened is you know the human one was already made, it was already posted, it got a lot of views. So I’m assuming that the skill here just took this clip because it knew this was like a good one. But it’s more like going through and saving the principles, you know, the theories of what makes this you know successful, quote unquote. So I see. Okay, gotcha. In this example, it analyzed 27 of our reels that had done really well. All of these were human edited, and it pretty much just watched them through like frame by frame. Here it’s just kind of breaking down the way that it’s able to look every couple seconds and understand like how often things change and and how the the viewer feels when you switch between this sort of B-roll. You’ve got like the shimmering effects, and it really it even got this granular.
[360s - 480s] It started to break down when my voice and the music and sound effects and the difference in the audio levels and it really, really just broke it down as granularly as it could. And it built me a bank of patterns. And then we took those patterns, turned it into a skill, and kept building on that and replicating it. So that is, you know, one example, right? So now I can drop in these raw videos, and I’m able to get outputs that look like they were edited by my human team, right? They’ve got the same same fonts, same vibe, same pacing, my style, pretty much. And obviously, it saved all this kind of stuff. Now, this is using you’re using Claude Code to do this. Correct. Yeah, Claude Code. It’s using Hyperframes as well in the back end, and then it will use the browser, so it will go grab any B roll it needs to, screen record, click on things, screenshots, and then of course it will generate any AI videos it needs to with like you know Key AI or Higgsfield or something like that. Yeah, and I’m I’m starting to hear more and more as well that you don’t even need like DaVinci Resolve or Premiere Pro or anything like that. It’s just does this just kind of do it all autonomously without pulling it into an editor like that? Totally. Yeah, I’ve also been starting to work in Blender into some of this stuff too. So now it’s it’s able to create some really cool 3D scenes for me as well with just essentially the similar idea. I’ve never opened Blender once. I don’t even know what the UI looks like, but I just know it creates 3D scenes, and then I just connect Claude Code to it, and then it just does it all. It’s it’s really really absurd. I’ve never learned Blender in my life, and I’m I’m pretty much the same. I’ve made more stuff in Blender in the last three months than I ever have in my entire life. That’s so funny. Yeah, it’s it’s really interesting though because you know at least me, it’s made me a lot more productive, but it’s also made me a lot lazier. You know, like I no longer want to go do these manual things that I used to just be able to do just fine in a couple minutes, but now it’s like I don’t want to do that, and I don’t know if that’s a good thing or a bad thing. I think you could argue for both sides, but yeah, I mean, I guess you know the argument is maybe your your sort of mental capacity has moved up the value chain a little bit, right? Like you are actually now.
[480s - 600s] More of the orchestrator thinking through the bigger picture, the bigger problems. Yeah, I agree with that. Then I started to do another one here for YouTube videos, pretty much from when I started doing YouTube up until now. Every one of my tutorials I’ve just edited myself, but I figured that it would be worth a shot, right? And for a while, one of the things I didn’t like about when I would have AI edit my videos was, you know, sometimes it would leave in two takes of something or just not catch some of the things that I catch. Like sometimes when I’m recording or editing a video, I’ll be like, hmm, I didn’t really explain that sentence very well. Let me actually just re-record this real quick. I’ll have the technical understanding of what’s going on in order to make sure that the the video makes sense. But now I’ve been able to work in some of that into the skill because now I make sure that it’s not only transcribing things and you know cutting it up, but it’s also thinking about the viewer’s perspective and and where there might be gaps and it’s surfacing some of that stuff to me, which is pretty cool. But the main thing I wanted to focus on here was pretty much the exact same framework, right? Like I took multiple videos that I edited myself, and now I can give Claude code my face cam recording as well as the screen recording, and it will place it in this this scene for me. And I pretty much have two scenes when I make a YouTube long form. It’s this scene where you can see like my face cam screen and then kind of this background, or just like my full screen face cam. So it also understands, okay, when has the screen not moved for a little bit? You know, Nate’s mouse has been still, and maybe Nate’s going on a rant or he’s talking about something, and that indicates for me to go to the full screen rather than just keeping you know this this view there. And then of course, what’s really nice is it will do certain zoom ins that I wouldn’t have done myself, and even something as little as this, right? Like just the little zoom in that I wouldn’t want to do myself. It has the visual understanding to know that a viewer is probably going to enjoy this better because now they can read the text a little bit more. Yeah. So it’s not just going through and very simply cutting out mistakes or cutting out the dead space, but it’s also really thinking about the experience of the viewer too. And I think that’s that’s one of the tougher things to work into these skills, but it’s an investment. Every time you give it feedback, it gets better. And then also, I’m
[600s - 720s] Able to work in essentially sort of like LLM as a judge, so that I can build up a better judge over time, so that the skill does keep improving. So Claude Code obviously it’s not capable really of watching videos, right? So how is it when you feed it a whole bunch of your old videos? How is it figuring out what you’re doing in those old videos? Is it just like like taking screenshots at certain intervals and and looking at those to figure it out? Because I think as far as I know, like Gemini is really one of the only ones that can actually watch videos right now. Yeah, that’s that’s exactly what I think as well. I did see there’s like this repo that maybe lets Claude Code do it, but I mean ultimately the way I think about it is yeah, just like let it screenshot basically every frame for YouTube videos. I didn’t really make it do that for these these short form ones. I certainly did. I think the first time that I really started to do this whole screenshot loop was when I was doing a lot of web design with Claude Code. You know, I would give it a reference example, or I would you know there was lots of things that were out of bounds, and I would just bake in. You have to screenshot every single element, like you scroll down, you screenshot everything, and then you look at it and make sure that things are in bounds. And I and then from there, I basically just baked in screenshotting into like everything that I do whenever it’s anything that you know takes a little bit of visual skill from the agent. So yeah, gotcha. And then I’m imagining it’s also using some sort of transcription API or MCP or something in the background that transcribes it, so it knows where to put each animation and that sort of thing. Yes, 100. Let me actually pull this up real quick. So what I just pulled up right here is I just asked it the other day to make me a video about how it edits some of my videos because at the end of the day, you know, like I’m just telling it all the stuff I want, and sometimes it’s like I don’t know exactly how you did that, but I like it. And it broke down. You know, I’ve got tons of different editing skills. I’ve got raw reels, which is more of like this style where I’ve got my you know I just whip out my Osmo and I just record my screen right or I flip it around and record me talking. I do like motion show reels. I’ve got the YouTube Shorts, the Instagram reels, curriculum videos. So anyways.
[720s - 840s] It’s just really, really cool what we’re able to push out with all these skills. But you were asking about, you know, the transcription and everything like that, and this is one of the big pieces of it that I had to work in. So obviously, it would listen, right? So it would transcribe the audio. I use Eleven Labs API to do it, just because it’s really quick. It’s super cheap too. I mean, you could also do something free like the Whisper, and it’s local, but it just it’s slower. So I use the Eleven Labs API, and it gets the words. And what it does is it will tell you, you know, like this word played from, or you spoke this word from point zero nine to point one seven, and that way you can actually sync the motion graphics to come in exactly when the words come in. So that’s very important. But then it would still sometimes clip words off, like halfway through the end of, like I would say videos, and then it would just come out as like video, like it would just cut me off. So rather than also just listening and syncing to the words, it would also look at the I don’t know what you call this, like a waveform. What is this? Yeah, whatever this is called, waveform, yeah. And it would make sure to do it when the words actually over by looking at when this hits like flat. So it used to just kind of cut here, but now I made sure the skill cuts here instead. It understands pointing, right? Like it understands that if I point here to make the claw logo go there, and if I pointed there earlier in this video, and the Grokbot logo goes there, so it’s just really, really intelligent with that sort of stuff. This is awesome. So what one final question on this one, and then we could sort of move on to the next workflow. But if somebody wanted to to sort of follow this, is there any place where they can find maybe your skills that you’ve already created, or how would somebody go and replicate what we’re talking about here themselves? So I’ve made a few videos on my YouTube channel about all of this stuff, but I’ve also got a GitHub repo which is public and it has pretty much all of these skills inside. So you could you could clone that in and then just you know ask about which skills would fit your workflow, and you should be able to you know be in a pretty nice spot with it. All right. So I know not everyone has a need for editing videos, although with it being that easy these days, I don’t know why most people wouldn’t start adding videos.
[840s - 960s] The mix. However, this next thing we’re going to talk about is something that I think everyone can use, because sometimes maybe you have an idea that you don’t know if you should pursue or not. Well, this next workflow that Nate’s going to show me is called the roast room, and it puts a panel of the smartest people in the world in front of you to debate about your idea. All right, check this out. So this one is called the roast room, and then I, you know, I was building off these different councils and things like that. But this is just a visual interface on top of what actually happens inside of Claude Code. So really, the idea here is that we have a council of different agents, and you’re able to ask for advice on some sort of, you know, whether it be personal or a business decision. Most of these people are going to be, you know, giving you like more more business focused advice. You pretty much just shoot in a prompt. They will all take an opening position. They will then start to argue with each other, and we can see all that happening. And then they’ll vote, and they’ll give you basically a final option with a rundown. And then of course you can look through the actual debate itself, and you can dig in a little deeper, and maybe even do another round. It really helps you look at it from different lenses, and especially if you start to mix and match, because I’ve also got some people in this roast room that aren’t currently on this council, and I’m adding more people to it. And then you can see there’s a skill in here called, I think it’s called new roaster. And so if you want to add your own roaster or council member, you just run this skill and then say who you want, and it will create them inside of this as well. So so it’s going to go and like research them and figure out sort of what they’re known for, what the their sort of philosophies on things are, stuff like that, and and pull it in as context. Absolutely, yeah. And maybe before we run a demo, I can show you like what actually is inside this, because I think it’s kind of cool to see, like like Alex Hormozi, like what even is that? What what is he in this context? So let me pull that up. We can quickly walk through that, and then we’ll run an example of the roast room. Okay, so this is again just a GitHub repo that anyone can grab, but you can see that we’ve also got down here we’ve got Naval, Codie Sanchez, Leila Hormozi.
[960s - 1080s] Sara Blakely and I’m probably just going to keep adding more people. I’m surprised I don’t have Elon Musk on here. Should probably have him. Should probably have like Elon Musk is just in there to troll everybody. You can see in here if I go to the dot Claude, we have these different agents. So if I open up the agents, this is basically what they are. They’re sub agents and they get triggered. Now one of the questions I get about this is why not an agent team? And I actually used to run this as an agent team. The roast skill used to be an agent team, but I ended up basically finding that it was costing me more for minimal gain in the actual values providing. So I switch it back to sub agents. But you certainly could set this up as agent teams if you want. But anyways, let’s just real quick open up Hormozi for example. This is his markdown file, which is just a very simple sub agent, right? We can see that you hold offers, pricing, and sales. So each of the council members basically have one question that they have to weigh against, which in this case for Alex is will people pay, how much, and how do we close? So very very Hormozi coded. But anyways, you can see that we have a brain, which I’ll look at in a sec. We have his lens, so the way he thinks, what he looks at, how he talks, how he debates, and things that he never does essentially. So each of these council members is just their own agent .md file, a sub agent file. And they’re all running the same model. They are, yes, they all run the same model, and you could customize that if you want, but it just runs the default. So if I go to the skills and we open up the roast room, you can see that there’s brains, references, and scripts. Really, the brains is the main one that we want to pay attention to here. So each of the agents has their own brain. So here you can see for who is this open? Gary V. His brain lives right here inside of the skills inside of his Gary brain. So I’ll open up his for example. Each of these people’s brains has these different markdown files. We have hot, which is basically his hot seat. So the question he owns, his worldview in three sentences, and then they all basically have eight rules they decide by. So this is all pulled from real sources. So
[1080s - 1200s] For Hormozi, for example, is one of the only ones I really remember off top of my head. We looked through like 123 different videos and book frameworks, articles, X posts, LinkedIn posts, and it was over like 1.5 million words that came straight from Hormozi. And then we build all these themes. And there’s also like specific quotes. So we have this hot. We have books. So this will go through the books and different frameworks and theses from each one. We have frameworks. We have positions. So yeah, I think you you understand the point here. It just does a bunch of research and then it organizes it in this nice way where we can look at these main concepts and even like stories. So it’ll pull in specific stories that they use. And there’s a lot of data here. This is all the context. So when you ask a question to the the panel that you’re asking to roast you, when they each individually respond, this is their sort of worldview, right? This is the context they have of with how they’re going to respond to the the questions. Exactly. Yep. They pretty much will look through all these and apply this on top of the question that you asked. So yeah, 100%. Gotcha. Very cool. Yeah, and then we’ve got the obviously the new roaster skill right here, which basically just helps you turn any person that you want to into one of these council members. So it’ll create even like the little like the cartoon image of them. So yeah, that’s a little a little peek under the hood of what this actually is. So I think that helps kind of you know before we we jump into an example. Now with the the actual visuals of this, did Claude code code up all the visuals? Like, are those graphics? Obviously, Claude code doesn’t have an image generator, but how was that that created? Yeah. So it I’m pretty sure it used GPT image too. It shot a request off, but obviously it used like their references. It originally some of these people did not look like how they really look at all, but now they all have the same sort of like cartoon look, and you can see Hormozi’s got like the classic acquisition cap, and yeah, so that’s all baked into the skill, and now you can add different council members, and it should also ask us based on the question, it should recommend like, hey, maybe these five would be best.
[1200s - 1320s] Or maybe you know these eight. So Matt, I guess I’d love to put you on the spot here. What do you want to ask the roast room? Is being a YouTuber a good business model? Yeah, I mean, we certainly can throw that out there. We have to just do one thing though. We have to give options, so they will actually vote on different options that we present. So just like if you have one idea, if you’ve got two different options you’re thinking about, or maybe five options you’re thinking about, just kind of throwing out at least some context of what you’re thinking. One thing that’s been an interesting discussion internally with me and my team lately has been our newsletter is going to continue to be worth it in the future. So I’m trying to think of like what sort of options to give, but that is that has been a constant narrative that’s been going on internally because we’ve got a very large newsletter, we’ve got 260,000 subscribers for like you know AI news. I am a little concerned that over time AI agents are just going to be the ones reading newsletters. There’s not going to actually be human eyeballs on there. So that’s something I’ve been thinking about a lot lately. I’m just trying to think of what the the sort of options to give would be. We could shoot off this context right. So Matt has an AI newsletter. It’s got 260,000 subscribers. He is concerned about agents reading these instead of humans. How how is your newsletter like being monetized? I guess that piece of context might be helpful as well. So right now it’s monetized through purely sponsorships. Okay, well they might be able to also create options for us and then vote on them. So I’m just going to say like here’s some context about a situation that Matt is thinking about with his team. Could you just like turn this into a roast room prompt for us and then kick it off? So let’s try that. Maybe it’ll just throw together some stuff for us. We’ll see if it wants to change up the council members. But yeah, it’s going to write that prompt, run the council, and then we’ll get to see it in real time. It looks like they actually shot this off, and you can see that they gave us five options here. Keep doing what works now. Rebuild for humans. Resell sponsorships on outcomes. Design for agents. Sell agent reach and shift effort to other channels. So what’s happening now is they’re.
[1320s - 1440s] Going to go through and make their opening positions. Oh, you can also see here it decided to take out Dan Martell and Gino, and it added in Naval. Oh, interesting. And surprises me a little bit. Yeah, I guess his main question here is who does this and what does it cost the founders’ calendar. That’s the one that he looks to protect. Naval is does this build leverage and equity you own, or does it just rent out your time? Yeah. Okay. Yeah, that makes sense then. So this one, like I said, round one they’ll take their positions, and then round two is when we actually see them like pick out who they want to like debate with or poke holes in their plan or whatever it may be. Now, is there actually audio on this when they speak it out, or is it all text? There’s not. There’s not. Someone said that I should add that and try to get their voices like a ElevenLabs or something and have them speak. That’d be pretty funny, but no, there is not audio. It’s just a weird little bobblehead-looking mouth movement. So, but now you can see that they made their initial initial picks, right? So we can see that Rebuild for Humans is one of the main is the winner. I don’t exactly know what that means, but I guess we’ll find out. But now they’re going to go through, and we’ll see them kind of decide who they want to chat with. And so, will they bring in different people, or is it it stays the same panel at this point? At this point, it’ll stay these same six council members. Gotcha. And we’ll just see them sort of make connections. But of course, if this local host wasn’t open, this would just be running right here. You know, like obviously, it’d be doing the exact same thing. It would create the same deliverables, same sort of vibe. But this is just way more fun. So. Yeah. Yeah. So everything’s happening basically in in the sort of AI chat. The visual is just to to make it a little bit more interesting and entertaining to watch. 100%. Yep. There we go. So now we can see that Gary V was going after Hormozi a little bit, and when they do stuff like this, if I scroll down here, we can actually watch like their opening positions. We can watch the key points. We can see the debate. So Gary V versus Hormozi. He said that he disagrees with Hermosy saying the readers are fine, and the broken part is a sponsor offer price on opens. I can see it here. Here are some other.
[1440s - 1560s] Things here’s my scoring, and so we can go ahead and read all of the actual like debates and discussions that are going on. But now I just wanted to scroll back up so we can see, you know, like Paul Graham also disagreed with Hormozi, and it looks like Bezos disagreed with Warren. So once they feel good about this debate, they’ll kind of like reset once again, and then we’ll see them actually do the vote, and then maybe these scores will have changed a little bit. Maybe they won’t have changed that much. But that’s what you can see next to their names. Yeah. So you can actually see that everyone voted B except for Warren and Alex. They both voted C. So are the are multiple debates happening at the same time? Because you’re seeing all those lines between different people. So those like they’re happening happening simultaneously. They pretty much are just happening one after one at a time because then the council can look at what these other two people discussed. Okay. Gotcha. So now we’re going through round three. We’re going to see them actually like vote, vote, vote, vote, vote, and then we’ll have a little PDF that we can look at, which is a very, very high level snapshot. But then of course, if we wanted to dig deeper, ask more questions, run it again, even we have everything saved, obviously in the logs. So what does the amend mean over their faces right now? It’s basically just when they come through and they have voted. There’s basically like one proposal, and then they’re either going to agree with that or they’re going to push back. So it looks like they they’re all starting to come to a place of agreement, and now they’re writing the verdict here. But yeah, if we wanted to really get in here and look at every single round of debate or discussion, then it can get pretty granular in here. Nice. And how brutal are they to each other? Looks like it’s pretty cordial. Yeah, that’s why we got to get Elon in the mix, right? Yeah. There we go. So we have go with B and run C beside it, which is interesting. So a summary passes along facts, not Matt’s judgment. B is the only option that tests whether readers want Matt pricing on opens prices on the number against agents eat about 78k a year is at risk on assumed numbers and click terms protected at no cash cost. So we can.
[1560s - 1680s] Actually, open up this PDF real quick, and that will just, like I said, it will give you the the super, however, overview. So this is what that looks like. It’s pretty much the exact same thing. It’ll show you the scoring, the final call. It’ll show you how they got here and like a finished round scoring of all of these different options with their weights and things to watch out for. And then it kind of gives you like an action plan. So yeah, just a super simple breakdown of what was actually discussed on the council. And then you know the probably the easiest way to set something like this up would be to go into like your Claude code, point it at the GitHub repo that you know Nate made available for everybody, and just say install this for me, right? Like that’s almost the easiest way to do any of this kind of stuff these days. Yep, literally just dropping the URL, set this up, and then you’re you should be good to go. Yeah. Awesome. We’re not even done yet. Nate is a master at finding killer ways to use AI, so I wanted to squeeze just one more idea out of him. So this time we’re going to tackle social media. In fact, we’re going to put Instagram on complete autopilot. This next workflow shows you exactly how Nate is putting out high quality content on Instagram using AI and is practically hands off technique. Check this out. So what we’re looking at here is it’s a very humble Instagram account. I’d say this has been going for about a month. At this point, ninety nine percent automated, maybe ninety eight. The goal is to get this a hundred percent automated, and then you know be able to monetize through sponsorships and things like that. But really, just to have a nice automated Instagram account that’s dropping news. So before I get into the workflow, we’ll just give a quick overview of what actually happens here because there are some things in here that are pretty cool. So some of these obviously are just you know grabbing a cool video that may be trending on X and just kind of throwing it in here, right? We’re crediting the original person, and then we’re just kind of ending with like a hey, yeah, moves fast, so go ahead and follow, right? Like nothing too complicated. Obviously, we had to build a skill around this so that we have a few different scenes, and that everything is very branded. But what else is cool is if we look at a different one like this one, for example.
[1680s - 1800s] The opening carousel was an AI generated image. We’ve got an image of Dario crying with a bunch of money, and the idea is that I’m able to either drop in an X post that I find, right? I can just literally send Claude the link, and then it just makes a post. So that means that it goes and it does research. It’ll go to a thread. It will pull off different reactions. It will also go like read articles, and it will either create B roll or it will go to the article. It even found this clip, right? It found a clip of Dario, threw it in here with audio. So it’s more than just making a carousel. It’s it’s automated like the backend research and thought that goes into making the carousel as well. And I’ll just go through one more thing that this does. So this one, for example, did really well, and you’ll notice that it has 559 comments. And this is obviously using an automation, right? Like a ManyChat type of automation, but I’ve automated that too. So we’re giving away a resource guide, and we could link people back to our community if we wanted to. I haven’t done that yet here. Just giving away the stuff, but not only are they commenting for it, which helps like boost engagement and things like that, but in the DMs with these people, we’re making them follow first to get the automation or to get the prompt. So they’re commenting, and then it says, “Hey, if you you know drop us a follow, we’ll send it over, and then we shoot it over. " And so all of that’s automated, and now, like I said, it’s you know 99 because every four hours it goes through, it finds a few posts, and it makes some carousels and posts them. But then I’m also still, if I see something, I’m still dropping it in, and I am you know maybe sometimes still iterating on the prompts a little bit. Sometimes I’ll see a post that I didn’t like, I’ll delete it, I’ll shoot it back and say, “Hey, you know, I didn’t like this piece of it. Do it again, blah blah blah. " But yeah, I mean, this really will get to a point where it is 100% automated. You know, this previously would have probably been a full time role for someone to manage and run this Instagram account. And once we do get to monetize it, I will 100% have agents that are managing the sponsorship inbox, automatically taking payments, automatically creating those posts, scheduling them out, communications. That is going to be super super cool, and you know I think that just goes to show 30 days, almost 2k followers. 2k is not that impressive, but I think it’s a lot more realistic of what is possible as long as you’re iterating on.
[1800s - 1920s] Impressive for thirty days, honestly. I mean, you know, if you told me you’ve been working on this for nine months and you’re at two k, that’d be one story. But thirty days in, I I think that’s that’s great for thirty days. Yeah, no, I’m I’m pretty happy with it, especially if you if you really think about like in total, the past thirty days, how many total hours have I spent on this? I would probably tell you four to five. Yeah, I’m pretty happy with it. Now, what I thought would be we could do real quick to end this episode or to end this demo would be I could go grab a quick X post and just shoot it in, and we could see it run. So I guess you can see this. Here’s an example I just found. AI could become self aware. We’ve got a video here, and what I can do is bring this over to Claude. And so what I always do here is, you know, you can see that I I got out a different one. Now I just post it, and it posted it. But then what I always do is I keep this on remote control, and then I keep my computer awake so that if I’m like at the gym or on a walk and I see an X post, I can just shoot it off. But also I have this connected to like a VPS version of Claude code, so I can do it there as well. But I basically would just say make a carousel for this, shoot the link off, and then when I run that, it goes there. It’s going to read the replies. It’s going to you know download the video, slice it up, and it will pretty much just make the whole research post. It might decide that there’s a good resource to have people DM for. It might not, but obviously we can enforce that. You know, if we wanted that to happen every time, we could enforce things like that. But yeah, it’s it’s just pretty cool, like automated pipeline now of creating these social media posts. And you’ve got a whole skill built, so when you say make a carousel for this, it just goes and knows what skill to grab. Yep, and that skill has things like our brand guidelines. It has things like the way that we talk. It has assets it will save. So sometimes if it’s creating generated videos or images, it will save them too in case we want to reuse them later. In this one, it’s going to it took the clip already. You can see it’s going to transcribe it, and my guess would be that it’s going to cut that up and have a few like twenty second clips in the carousel of like important things that this person said in the clip. But we will see what it actually does. Here’s another example of like obviously it’s it’s taking screenshots and stuff, right? Because when it does things.
[1920s - 2040s] Like this, where it’s putting this little thing behind Dario. It has to like cut out Dario, put that in there, and then layer it back. And it has to make sure that it’s not going to cover up his face or anything like that. And sometimes it’ll just play a video here too. So it will make sure that it’s spacing everything out right, that it actually makes sense. How long does this typically take to run? So once you feed it a video and say, “Hey, make this carousel,” what would you say like the average amount of time it takes to do it is? I would say it’s probably about five minutes. Maybe this wasn’t the best example because whenever it’s a video, it obviously takes a little bit longer. But it doesn’t take too long. And then you can also just like I’ve come into a thread before and just dropped in like like like ten posts that I found that I liked, and then it will just bang all of them out. And also the desktop app, it does this cool thing now where it will suggest to open up another thread, and then it will just like slot in a thread below this one, and they’re all linked. And then you can have this one main thread manage all the other ones. I don’t know if you’ve played around with that yet, but it’s pretty slick. When uh when you shoot off a ton of work to one thread, it’ll be like, hey, you know, these are all kind of separate tasks. Let me shoot these off in like their own like parent or or child sessions or something. Yeah, yeah, that’s kind of like the new projects feature in Claude. Yes, yes, that that that’s what it is. Yeah, super. It is pretty nice. And also, that’s another piece that it does is it confirms everything. Oh, nice. So it makes sure that everything is real because sometimes on X there’s a lot of like fake news, right? And then the other thing it will do is if I’m manually asking it to make a carousel when I say post it, before it actually does it, it will go make sure everything is still true because sometimes I’ve had it happen where I have it generate a carousel and then I ask it to post it, and then the old post actually got deleted or edited or something else has changed. So it does like one one more sweep through, which is nice. Instagram’s API is kind of rough, right? So I’m imagining it’s actually opening the browser in Claude, and you’re already logged into Instagram, so it posts inside of the Claude browser that’s already logged in. Is that what’s happening, or is it going directly through the API? Actually, this one is using Botato everything. I’m not familiar with that one. It’s the tool that Sabrina Ramonov created. If you’re familiar with Sabrina.
[2040s - 2160s] But I’ve heard her name before. Yeah, so it’s it’s a tool that she created and it allows you to have multiple accounts. So, like I told you earlier, I’ve got a few other things that are auto posting on X and LinkedIn and Instagram, TikTok, bunch of different accounts. And Botato can just manage all those accounts. And then it also she just pretty much worked in the DM automation functionality. And that’s when I was like, you know what, I’m gonna I’m gonna test this out on you know this this account. So the DM automation stuff isn’t actually through ManyChat. It’s through what did you call it? Botato or? Botato. Botato. Okay. It’s the ManyChat style, but yeah, it’s not it’s not through ManyChat. It’s using Botato. So you can see this one finished up. It took I think it said a little over five minutes, like five or six. And what it does is it gives you this video, so you can actually just like preview the entire carousel. We’ve got the opening slide, and it chose to do this tweet style. So it played a lot of the video in just one slide. So I think this is a two-slide carousel, and it just ended it with that. So it was looking through to see what would be a good frame for the last picture. But yeah, then it just basically put this with these with this caption. These systems might acquire some feeling of self-awareness. You can see. Oh wait, there are three video tweet, thirty six seconds, and then another one’s thirty six seconds. Oh, okay. So halfway through the text switches to we’re made of the same carbon matter with our squishy brains, and then you can see here it’s already recommending to post it. So I would just hit tab, hit enter, and you know what? I might I might as well just go ahead and post it. But then it’ll do a quick refresh and make sure everything’s still good, and then it will just pop up on my my Instagram and say that’s live. And it also logs everything on a a Google Sheet, so I can go see every single post. And that way, the reason I’m having it log everything is just so that it makes sure that it’s not duplicating a topic. So I’ve also got that going on too. You know what would be cool, and this is thinking out loud, is if periodically it pulled in stats of how they’re performing into the Google Sheet, and then it can go look back at what stuff performed and almost have like its own internal reinforcement learning of how to get better and better because it’s watching the stats of what performed. Yes, you are you are 100 right, and I forget if that is automated or not, but here.
[2160s - 2280s] I’ll actually just show you real quick. Okay, so I haven’t set that up yet, but yeah, I’ve prepped it to do that so that over here you can see it will go through and it will start to come back and understand the stats on it. I guess it did it a weekly for some of these, but not all of them. But yeah, I do need to set up that element, and then that’ll be really cool to see how it’s able to do like a weekly analytics and see if it updates its own skills a little bit based on the data it’s finding. So, 100%. Yeah, yeah, sort of self-improving loop. That’d be super cool. All right, that’s what we got for you today: three workflows that you can go and implement right now. And just as a reminder, like the easiest way to go and implement these is go to the GitHub page where Nate made all of this stuff available. Copy in the link to that GitHub page to your Claude code or your Codex or your Grokbot or whatever app you’re using, and say, “Go build this out for me,” and it will just build everything for you. But once again, I want to thank Nate Herk so much for jumping on with me and breaking down his workflows and going into so much detail so that I could share them with you. And you have some really cool practical use cases for AI. This is a series that I’m going to be doing more of. This is my “Show Me Your Workflow” series. The last one with Bolavos Sudo showed off his God’s Eye View, and now we got three more workflows from Nate. If you like these types of videos, let me know in the comments, and I’ll make more of these “Show Me Your Workflow” videos. And if you haven’t already, maybe consider liking this video and subscribing to this channel because that will make sure that when future versions of this pop up, they’ll show up in your YouTube feed. And if you’re brand new to this channel, I also try to keep everybody super looped in with all of the AI news. Every Friday, I make a breakdown of all of the AI news that happened throughout the week, so that if you don’t want to have to drink from the fire hose all week and be overwhelmed by all the AI news, well, I make one video that you can check into once a week and get fully looped in for the week. So if that’s something that interests you as well, definitely subscribe to this channel, and I’ll make sure those show up in your YouTube.
[2280s - 2289s] YouTube feed. Thank you so much for hanging out with us today and nerding out with us. This was a blast to do, and hopefully, I’ll see you in the next one. Thanks so much. Bye bye.