YT YouTube2Text

YouTube2Text → Transcripts → Nate Herk Breaks Down His 3 Most Powerful AI Automations

Nate Herk Breaks Down His 3 Most Powerful AI Automations — Transcript

by Matt Wolfe · 9,035 words · 1,263 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00This channel is all about finding you
0:01the coolest AI tech, tools, models, and
0:04perhaps my favorite thing, which is
0:07workflows. This video is the second in a
0:09new series that I like to call "Show Me
0:11Your Workflow", where I bring guests and
0:13friends onto the channel and then I
0:15mercilessly squeeze every last drop of
0:18value that I can from them. And by
0:20squeeze, I mean I get on a Zoom call and
0:22politely ask them questions about the
0:24cool stuff they built. And today, I'm
0:26talking with Nate Hurt, one of my
0:28all-time favorite builders of cool
0:31stuff. He's going to walk you through
0:33three of his favorite AI builds that he
0:36uses every single day. His automated AI
0:39editing workflow, his AI expert panel
0:42called The Roast Room, and his AI social
0:45media manager. And the best thing is, by
0:47the end of this video, you'll be able to
0:49do all of this stuff on your own, too.
0:50So, Nate Hurt, show me your workflow.
0:55>> [music]
1:00>> Let's optimize that workflow.
1:03>> So, it sounds like the first one we're
1:05going to get into is like the AI editing
1:07workflow, which is This is something
1:09I've been hearing a lot about from a lot
1:10of people, and I'm really, really
1:12excited to hear your take on this
1:14because
1:15I've tried to do some just like let the
1:17AI edit for me, and I don't know, I
1:20always find that like I'm spending more
1:22time when I let it do it for me cuz I'm
1:24going back and sort of fixing a lot of
1:27stuff. So, I I'm really, really excited
1:29to hear your your editing workflow here.
1:31>> Yeah, this has been something that I've
1:34been very, very surprised by in a good
1:37way. Because I think when all this
1:39started, I agree with what you just
1:41said. It was just like it was
1:43frustrating because it would take more
1:44time because the rendering would take a
1:46while, and then you would make a a
1:48change, and then it would have to
1:49re-render everything. But, once I really
1:52started to change the way I thought
1:54about how I was doing these editing
1:55workflows, as well as, of course, saving
1:57all of the knowledge and the feedback
1:59and building skills around it, then this
2:01has truly made me so much more
2:02productive. Like it's it's it's pretty
2:04nuts. Like just this morning, for
2:05example, I had worked this in
2:08combination with some other skills that
2:09I'll talk about in a bit, but was able
2:11to just bang out like 20 reels. It did
2:13all the ideation, all the scripting, and
2:15then I just threw it on my teleprompter,
2:16read off. The cool thing is like it
2:18grabs all the B-roll and it makes all
2:19the AI-generated videos that go into it.
2:21So, it's it's really a cool pipeline.
2:23But, I wanted to start off here before
2:24maybe we look at like some some examples
2:26and stuff about how it was built. And
2:28this is kind of a wordy doc, but I just
2:30wanted to talk about the the way I think
2:32about not only this workflow, but in
2:34general the way that I build skills and
2:36and workflows that I think is really a
2:37game changer. So, it's kind of timely
2:39with the whole, you know, reverse
2:41engineer anything repo that dropped and
2:43just went insanely viral.
2:44>> Oh, yeah.
2:45>> The whole idea is that I've been trying
2:47to build skills around editing
2:49short-form content for a while, and it
2:50just was never really there. It just
2:52always felt a little cheap. And AI on
2:55its own, it's not very great at thinking
2:56outside the box. I mean, if you really
2:58think back to how it was actually like
2:59what it does under the hood, like it
3:01just predicts the next word, right? Like
3:02at the most fundamental highest level,
3:05which means that it's going off of like
3:06what most people would assume or like
3:08what makes sense. So, what I started to
3:10do was, you know, I had all of these
3:12reels that were proven. Now, I can take
3:14this raw input where I'm speaking,
3:16right? Like I'm just able to record even
3:17if I have mistakes in there. I'm just
3:18speaking. And then, what I have right
3:20here side by side is the manual versus
3:22the skill now. So, this is like a human
3:24editor had built this. And you can see
3:26this the styling is very consistent, the
3:28font, the way the motion graphics come
3:30in, the shimmers, the sound effects, and
3:31the music are are pretty similar too.
3:33And that is one of the areas where AI, I
3:35think, still has a lot of work to do.
3:38[snorts] Music and sound design. I think
3:39music and sound design is super super
3:41underrated in the experience of for a
3:43viewer. And right now, like a human is
3:45just going to do that a lot better. But,
3:47it is getting there. And I'm starting to
3:49get to this point now where my Herc edit
3:50kit is what I call it. It's saved a
3:52bunch of the sound effects I really like
3:54and sounds that I like and like music,
3:57trending audios. And I'm just constantly
3:59feeding in like when I would want it to
4:01use this one. You know what I mean? So
4:03it's definitely getting better. But the
4:05point I was trying to make here is you
4:06can kind of see there's like three main
4:07scenes, right? There's me on the bottom
4:09with B-roll up top. There is B-roll full
4:11screen, and then there's me full screen.
4:13And basically the way I was able to get
4:15to somewhere where I really liked this
4:16is I had it go do what was proven. And
4:19so before I I continue scrolling down
4:20here, I think that's just a really
4:21really important mindset to have when
4:23you're working with AI is when you're
4:25trying to have it help you with
4:26something, the best inspiration is is
4:28proven work. So like if you think about
4:30when you're deploying AI in production,
4:32what do you do is you run evals. And
4:34your evals test against the golden data
4:36set. So essentially like why would you
4:38not start with that kind of golden data
4:40set in mind when you're building? Cuz
4:41then you know what you're building
4:42towards. So whether it be an editing
4:44flow or whether it be even just like a
4:46business decision, if you go look at an
4:48example of someone who's done exactly
4:50what you want to do successfully and has
4:52proven it, then break down how it works,
4:54how they got there, and I think that
4:56just if you zoom out and think about
4:57life in that way. Like it's changed the
4:59way that I I make decisions, but
5:01>> So real quick back up on the one that
5:03you were
5:04you were showing there. You've got the
5:06two videos side by side, and it looked
5:07like it was finding pretty much the same
5:10B-roll, right? Like it's the same video
5:12of Carpathi there. So the human one
5:14found the same B-roll as the skill, or
5:17did you tell the skill which B-roll to
5:19use?
5:19>> Well, in this case what happened is, you
5:21know, the human one was already made. It
5:22was already posted. It got a lot of
5:23views. So I'm assuming that the skill
5:25here just took this clip because it knew
5:27this was like a good one. But it's more
5:29like going through and saving the
5:32principles. You know, the the theories
5:34of what makes this, you know,
5:35successful, quote unquote. So
5:37>> I see. Okay, got you.
5:38>> In this example, it analyzed 27 of our
5:41reels that had done really well. All of
5:42these were human edited. And it pretty
5:44much just watched them through like
5:45frame by frame. Here it's just kind of
5:47breaking down the way that it's able to
5:48look every couple seconds and understand
5:50like how often things change and and how
5:53the the viewer feels when you switch
5:55between this sort of B-roll. You've got
5:56like the shimmering effects. And it
5:58really it even got this granular. It
6:00started to break down when my voice and
6:02the music and sound effects and the
6:04difference in the audio levels, and it
6:06really, really just broke it down as
6:07granularly as it could, and it built me
6:10a bank of patterns. And then we took
6:12those patterns, turned it into a skill,
6:14and kept building on that and and
6:16replicating it. So, that is, you know,
6:18one example, right? So, now I can drop
6:20in these raw videos, and I'm able to get
6:23outputs that look like they were edited
6:25by my human team, right? They've got the
6:27same same fonts, same vibe, same pacing,
6:30my style, pretty much. And and obviously
6:32it saved all the kind of stuff.
6:33>> Now, this is using You're using Claude
6:35code to do this?
6:36>> Correct, yeah. Claude code, it's using
6:38hyperframes as well on the back end, and
6:39then it it will use the browser. So, it
6:41will go grab any B-roll it needs to,
6:44screen record, click on things,
6:46screenshots, and then of course it will
6:48generate any AI videos it needs to with
6:50like, you know, key.ai or hix.ai or
6:52something like that.
6:53>> Yeah, and I'm I'm starting to hear more
6:55and more as well that you don't even
6:57need like DaVinci Resolve or Premiere
6:59Pro or anything like that. It's just
7:01Does this just kind of do it all
7:02autonomously without pulling it into an
7:04editor like that?
7:05>> Totally, yeah. I've also been starting
7:07to work in Blender into some of this
7:09stuff, too. So, now it's it's able to
7:11create some really cool 3D scenes for me
7:12as well with just essentially the
7:14similar idea. I've never opened Blender
7:16once. I don't even know what the UI
7:17looks like, but I just know oh, it it
7:19creates 3D scenes, and then I just
7:20connect Claude code to it, and then it
7:21just does it all. It's It's really,
7:23really absurd.
7:24>> I I've never learned Blender in my life,
7:26and I'm I'm pretty much the same. I've
7:28made more stuff in Blender in the last 3
7:30months than I ever have in my entire
7:32life.
7:33>> That's so funny, yeah. It's It's really
7:35interesting though, because, you know,
7:36at least me, it's made me a lot more
7:37productive, but it's also made me a lot
7:39lazier, you know? Like,
7:41>> [laughter]
7:41>> I no longer want to go do these manual
7:43things that I used to just be able to do
7:44just fine in a couple minutes, but now
7:46it's like, ah, I don't want to do that.
7:47And I I don't know if that's a good
7:47thing or a bad thing. I think you could
7:49argue for both sides, but
7:50>> Yeah, I mean, I I guess, you know, the
7:52argument is maybe your your sort of
7:55mental capacity has moved up the value
7:57chain a little bit, right? Like, you are
7:58actually now more of the orchestrator,
8:01thinking through the bigger picture, the
8:02bigger problems.
8:03>> Yeah, I agree with that. Then I started
8:06to do another one here for YouTube
8:08videos. Pretty much from when I started
8:10doing YouTube up until now, every one of
8:12my tutorials I've just edited myself,
8:14but I figured that it would be worth a
8:16shot, right? And for a while, one of the
8:18things I didn't like about when I would
8:19have AI edit my videos was, you know,
8:21sometimes it would leave in two takes of
8:23something or just not catch some of the
8:25things that I catch. Like sometimes when
8:26I'm recording or editing a video, I'll
8:28be like, "Hmm, I didn't really explain
8:29that sentence very well. Let me actually
8:31just re-record this real quick." I'll
8:33have the technical understanding of
8:34what's going on in order to make sure
8:36that the the video makes sense. But now
8:38I've been able to work in some of that
8:40into the skill because now I make sure
8:41that it's not only transcribing things
8:43and, you know, cutting it up, but it's
8:44also thinking about the viewer's
8:47perspective and and where there might be
8:48gaps and it's servicing some of that
8:49stuff to me, which is pretty cool. But
8:51the main thing I wanted to focus on here
8:52was pretty much the exact same
8:54framework, right? Like I took multiple
8:56videos that I edited myself, and now I
8:58can give Claude Code my face cam
9:00recording as well as the screen
9:02recording, and it will place it in this
9:04this scene for me. And I pretty much
9:06have two scenes when I make a YouTube
9:07long form. It's this scene where you can
9:09see like my face cam screen and then
9:12kind of this background or just like my
9:13full screen face cam. So it also
9:15understands, okay, when has the screen
9:17not moved for a little bit? You know,
9:19Nate's mouse has been still and maybe
9:20Nate's going on a rant or he's talking
9:22about something and that indicates for
9:24me to go to the full screen rather than
9:26just keeping, you know, this this view
9:28there. And then of course what's really
9:29nice is it will do certain zoom ins that
9:32I wouldn't have done myself. And even
9:34something as little as this, right? Like
9:35just the little zoom in that I wouldn't
9:36want to do myself, it has the visual
9:39understanding to know that a viewer is
9:41probably going to enjoy this better
9:42because now they can read the text a
9:43little bit more. So it's not just going
9:46through and very simply cutting out
9:48mistakes or cutting out the dead space,
9:49but it's also really thinking about the
9:51experience of the viewer, too, and I
9:53think that's that's one of the tougher
9:54things to work into these skills, but
9:56it's an investment every time you give
9:57it feedback it gets better. And then
9:59also I'm able to work in essentially
10:01sort of like LLM as a judge so that I
10:03can build up a better judge over time so
10:05that the skill does keep improving.
10:07>> So Claude Code, obviously, it's not
10:09capable really of watching videos,
10:12right? So, how is it when you feed it a
10:13whole bunch of your old videos, how is
10:16it figuring out what you're doing in
10:19those old videos? Is it just like like
10:22taking screenshots at certain intervals
10:24and and looking at those to figure it
10:26out? Cuz I think as far as I know, like
10:28Gemini is really one of the only ones
10:30that can actually watch videos right
10:32now.
10:32>> Yeah, that's that's exactly what I think
10:34as well. I did see there's like this
10:35repo that maybe lets Claude code do it,
10:37but I mean, ultimately, the way I think
10:39about it is, yeah, just like let it
10:40screenshot basically every frame. For
10:42YouTube videos, I didn't really make it
10:44do that. For these these short-form
10:45ones, I certainly did. I think the first
10:47time that I really started to do this
10:48whole screenshot loop was when I was
10:50doing a lot of web design with Claude
10:51code. You know, I would give it a
10:53reference example, or I would, you know,
10:54there was lots of things that were out
10:55of bounds, and I would just bake in, you
10:57have to screenshot every single element.
11:00Like you scroll down, you screenshot
11:01everything, and then you look at it and
11:03make sure that it things are in bounds.
11:04And I And then from there, I basically
11:06just baked in screenshotting into like
11:08everything that I do whenever it's
11:10anything that, you know, takes a little
11:11bit of visual skill from the agent. So,
11:13yeah.
11:14>> Got you. And then I'm imagining it's
11:16also using some sort of transcription
11:18API or MCP or something in the
11:20background that transcribes it, so it
11:22knows where to put each animation and
11:24that sort of thing.
11:25>> Yes, 100%. Let me actually pull this up
11:27real quick. So, what I just pulled up
11:29right here is I just asked it the other
11:31day to make me a video about how it
11:34edits some of my videos. Because at the
11:36end of the day, you know, like I'm just
11:37telling it all the stuff I want, and
11:39sometimes it's like, I don't know
11:40exactly how you did that, but I like it.
11:42And it broke down. You know, I've got a
11:43tons of different editing skills. I've
11:45got raw reels, which is more of like
11:48this style where I've got my, you know,
11:49I just whip out my Osmo, and I just
11:51record my screen, right? Or I flip it
11:52around and record me talking. I do like
11:54motion show reels. I've got the YouTube
11:56shorts, the Instagram reels, curriculum
11:58videos. So, anyways, it's just really,
12:00really cool what we're able to push out
12:01with all these skills, but you were
12:03asking about, you know, the
12:05transcription and everything like that.
12:07And this is one of the big pieces of it
12:08that I had to work in. So, obviously, it
12:11would listen, right? So, it would
12:11transcribe the audio. I use 11 Labs API
12:14to do it just because it's really quick.
12:16Um it's super cheap, too. I mean, you
12:17could also do something free like the
12:19Whisper, and it's local, but it just
12:21it's slower. So, I use the 11 Labs API,
12:23and it gets the words, and what it does
12:25is it will tell you, you know, like this
12:26word played from or you spoke this word
12:28from 0.09 to 0.17. And that way, you can
12:31actually sync the motion graphics to
12:33come in exactly when the words come in.
12:35So, that's very important. But then, it
12:36would still sometimes clip words off,
12:38like halfway through the end of like I
12:40would say videos, and then it would just
12:41come out as like video like it would
12:42just cut me off. So, rather than also
12:45just listening and syncing to the words,
12:47it would also look at the I I don't know
12:49what you call this, like a waveform?
12:50What is this
12:51Whatever this is called.
12:52>> yeah.
12:53>> And it would make sure to do it when the
12:55word's actually over by looking at when
12:57this hits like flat. So, it used to just
12:59kind of cut here, but now I made sure
13:01the skill cuts here instead. It
13:03understands pointing, right? Like it
13:05understands that if I point here to make
13:07the Claude logo go there. And if I I
13:08pointed there earlier in this video, and
13:09the Grok Bot logo goes there. So, it's
13:12just really, really intelligent with
13:13that sort of stuff.
13:15>> This is awesome. So, um what one final
13:17question on this one, and then we could
13:19sort of move on to the next workflow.
13:21But if somebody wanted to to sort of
13:23follow this, is there any place where
13:25they can find maybe your skills that
13:27you've already created, or how would
13:29somebody go and replicate what we're
13:31talking about here themselves?
13:32>> So, I've I've made a few videos on my
13:34YouTube channel about all of this stuff,
13:36but I've also got a GitHub repo, which
13:38is public, and it has pretty much all of
13:40these skills inside. So, you could you
13:42could clone that in, and then just, you
13:44know, ask about which skills would fit
13:46your workflow, and you should be able
13:47to, you know, be in a a pretty nice spot
13:49with it.
13:49>> All right. So, I know not everyone has a
13:52need for editing videos. Although, with
13:54it being that easy these days, I don't
13:57know why most people wouldn't start
13:59adding videos in the mix. However, this
14:01next thing we're going to talk about is
14:02something that I think everyone can use.
14:04Because sometimes, maybe you have an
14:07idea that you don't know if you should
14:09pursue or not. Well, this next workflow
14:11that Nate's going to show me is called
14:13the roast room, and it puts a panel of
14:16the smartest people in the world in
14:18front of you to debate about your idea.
14:21All right, check this out.
14:22>> So, this one is called the roast room.
14:25And then I you know, I was building off
14:27these different councils and things like
14:28that. But this is just a visual
14:30interface on top of what actually
14:31happens inside of Cloud Code. So really
14:33the idea here is that we have a council
14:36of different agents. And you're able to
14:39ask for advice on some sort of you know,
14:42whether it be personal or business
14:43decision. Most of these people are going
14:46to be you know, giving you like more
14:47more business focused advice. You pretty
14:50much just shoot in a prompt. They will
14:52all take an opening position. They will
14:54then start to argue with each other and
14:55we can see all that happening. And then
14:57they'll vote and they'll give you
14:58basically a final option with a rundown.
15:00And then of course you can look through
15:01the actual debate itself and you can dig
15:03in a little deeper and maybe even do
15:04another round. It really helps you look
15:06at it from different lenses. And
15:08especially if you start to mix and
15:09match. Cuz I've also got some people in
15:10this roast room that aren't currently on
15:12this council and I'm adding more people
15:14to it and then you can see there's a
15:15skill in here called I think it's called
15:17new roster. And so if you want to add
15:19your own roster or council member, you
15:21just run this skill and then say who you
15:23want and it will create them inside of
15:25this as well.
15:26>> So it's going to go and like research
15:28them and figure out sort of what they're
15:30known for, what the their sort of
15:32philosophies on things are, stuff like
15:34that and and pull it in as context.
15:37>> Absolutely. Yeah. And maybe before we
15:38run a demo, I can show you like what
15:41actually is inside this. Because I think
15:43it's kind of cool to see like like Alex
15:45Hormozi like what even is that? What
15:46what is he in this context? So let me
15:48pull that up. We can quickly walk
15:50through that and then we'll run an
15:50example of the roast room. Okay. So this
15:52is again just a GitHub repo that anyone
15:54can grab. But you can see that we've
15:57also got down here we've got Naval, Cody
15:58Sanchez, Leila Hormozi, Sara Blakely.
16:01And I'm probably just going to keep
16:02adding more people. I'm surprised I
16:03don't have Elon Musk on here. I should
16:04probably have him. Should probably have
16:06like um
16:07>> Elon Musk is just in there to troll
16:08everybody.
16:09>> [laughter]
16:09>> You can see in here if I go to the dot
16:11cloud, we have these different agents.
16:13So if I open up the agents, this is
16:15basically what they are. They're sub
16:16agents and they get triggered. Now one
16:19of the questions I get about this is why
16:21not an agent team? And I actually used
16:22to run this as an agent team. This the
16:24road skill used to be an agent team, but
16:26I ended up basically finding that it was
16:29costing me more for minimal gain in the
16:32actual values providing, so I switched
16:34it back to subagents. But you certainly
16:35could set this up as agent teams if you
16:37want. But anyways, let's just real quick
16:39open up Hermosi for example. This is his
16:42markdown file, which is just a very
16:43simple subagent, right? We can see that
16:46you hold offers, pricing and sales. So
16:48each of the council members basically
16:50have one question that they have to
16:52weigh against, which in this case for
16:54Alex is will people pay, how much, and
16:56how do we close? So very, very Hermosi
16:59coded. But anyways, you can see that we
17:01have a brain, which I'll look at in a
17:02sec. We have his lens, so the way he
17:05thinks, what he looks at, how he talks,
17:07how he debates, and things that he never
17:09does, essentially. So each of these
17:11council members is just their own
17:13agent.md
17:15file, a subagent file.
17:16>> And they're all running the same model?
17:18>> They are, yes. They all are running the
17:19same model and you could customize that
17:21if you want, but it just runs the
17:22default. So if I go to the skills and we
17:26open up the roster room, you can see
17:28that there's brains, references, and
17:30scripts. Really the brains is the main
17:32one that we want to pay attention to
17:33here. So each of the agents has their
17:35own brain. So here you can see for uh
17:38who is this open? Gary V. His brain
17:40lives right here inside of the skills,
17:41inside of his Gary um brain. So I'll
17:44open up his for example. Each of these
17:46people's brains has these different
17:48markdown files. We have hot, which is
17:50basically his hot seat, so the question
17:52he owns, his world view in three
17:54sentences, and then they all basically
17:55have eight rules they decide by. So this
17:58is all pulled from real sources. So for
18:00Hermosi for example is one of the only
18:02ones I really remember off top of my
18:03head. We looked through like 123
18:05different videos and book frameworks,
18:08articles, X posts, LinkedIn posts, and
18:11it was over like 1.5 million words that
18:13came straight from Hermosi, and then we
18:15build all these themes. And there's also
18:16like specific quotes. So we have this
18:18hot, we have books, so this will go
18:21through the books and different
18:22frameworks and thesis from each one.
18:24We have frameworks. We have positions.
18:27So, yeah, I think you you understand the
18:28point here. It just does a bunch of
18:30research and then it organizes it in
18:31this nice way where we can look at these
18:33main concepts and even like stories. So,
18:35it'll pull in specific stories that they
18:37use and there's a lot of data here.
18:39>> This is all the context. So, when you
18:40ask a question to the
18:43uh the panel that that you're asking to
18:45roast you, when they each individually
18:48respond, this is their sort of
18:49worldview, right? This is the context
18:51they have of with how they're going to
18:53respond to the the questions.
18:55>> Exactly. Yep. They pretty much will look
18:57through all these and apply this on top
19:00of the question that you asked. So,
19:01yeah, 100%.
19:02>> Got you. Very cool.
19:04>> Yeah, and then we've got the obviously
19:05the new roaster skill right here, which
19:07basically just helps you turn any person
19:09that you want to into one of these
19:11council members. So, it'll create even
19:13like the little like the cartoon image
19:15of them. So, yeah, that's a little a
19:16little peek under the hood of what this
19:19actually is. So, I think that helps kind
19:21of, you know, before we we jump into an
19:22example.
19:23>> Now, with the the actual visuals of
19:25this, did Claude Code code up all the
19:28visuals? Like are those graphics?
19:29Obviously, Claude Code doesn't have an
19:31image generator, but how was that that
19:33created?
19:34>> Yeah, so it I'm pretty sure it used GPT
19:37image too. It shot a request off, but
19:39obviously it used like their references.
19:41It originally some of these people did
19:43not look like how they really look at
19:44all, but now they all have the same sort
19:46of like cartoon look and you can see
19:47Hermosi's got like the classic
19:49acquisition cap and yeah, so that's all
19:52baked into the skill and now you can add
19:54different council members and it should
19:55also ask us based on the question. It
19:57should recommend like, "Hey, maybe these
19:59five would be best or maybe, you know,
20:00these eight." So, Matt, I guess I I'd
20:02love to put you on the spot here. What
20:04do you want to ask the roaster?
20:06>> Is being a a YouTuber a good business
20:07model?
20:08>> [laughter]
20:09>> Yeah, I mean, we certainly can throw
20:10that out there. We have to just do one
20:11thing though. We have to give options.
20:14So, they will actually vote on different
20:15options that we present. So, just like
20:17if you have one idea, if you've got two
20:19different options you're thinking about
20:20or maybe five options you're thinking
20:21about, just kind of throwing out at
20:23least some context
20:25of what you're thinking.
20:28>> been an interesting discussion
20:29internally with me and my team lately
20:32has been are newsletters going to
20:34continue to be worth it in the future.
20:36So I'm trying to think of like what sort
20:38of options to give, but that is that has
20:40been a a constant narrative that's been
20:43going on internally cuz we've got a very
20:44large newsletter. We've got 260,000
20:47subscribers for like, you know, AI news.
20:50I am a little concerned that over time
20:52AI agents are just going to be the ones
20:54reading newsletters. There's not going
20:55to actually be human eyeballs on there.
20:58So that's something I've been thinking
20:59about a lot lately. I'm just trying to
21:01think of what the the sort of options to
21:03give would be.
21:04>> We could shoot off this context, right?
21:06So Matt has an AI newsletter. It's got
21:09260,000 subscribers. He is concerned
21:12about agents reading these instead of
21:14humans. How how is your newsletter like
21:16being monetized? I guess that piece of
21:17context might be helpful as well.
21:19>> So right now it's monetized through
21:20purely sponsorships.
21:22>> Okay. Well, they might be able to also
21:24create options for us and then vote on
21:26them. So I'm just going to say like
21:27here's some context about a situation
21:28that Matt is thinking about with his
21:30team. Could you just like turn this into
21:32a roast room prompt for us and then kick
21:35it off?
21:35>> Cool, let's try that.
21:36>> Maybe it'll just throw together some
21:38stuff for us. We'll see if it wants to
21:41change up the council members. But yeah,
21:43it's going to write that prompt, run the
21:44council, and then we'll get to see it in
21:46real time. It looks like they actually
21:47shot this off and you can see that they
21:49gave us five options here. Keep doing
21:51what works now, rebuild for humans,
21:53resell sponsorships on outcomes, design
21:55for agents, sell agent reach, and shift
21:57efforts to other channels. So what's
21:59happening now is they're all going to go
22:00through and make their opening
22:01positions. Oh, you can also see here, it
22:03decided to take out Dan Martell and Gino
22:04and it added in Naval.
22:06>> Oh, interesting.
22:07>> And it surprises me a little bit.
22:09>> [laughter]
22:09>> Yeah, I guess his main question here is
22:11who does this and what does it cost the
22:13Founders Calendar? That's the one that
22:14he looks to protect. Naval is does this
22:17build leverage and equity you own or
22:18does it just rent out your time?
22:20>> Yeah, okay. Yeah, that makes sense then.
22:22>> So this one, like I said, round one
22:23they'll take their positions and then
22:24round two is when we actually see them
22:26like pick out who they want to like
22:29debate with or poke holes in their plan
22:31or whatever it may be.
22:32>> Now is there actually audio on this when
22:35they speak it out or is it all just
22:36>> [laughter]
22:36>> text?
22:37>> There's not. There's not. Someone said
22:38that I should add that and try to get
22:40their voices like a 11 Labs or something
22:42and have them speak. That'd be pretty
22:43funny, but no, there is not audio. It's
22:46just a weird little bobblehead looking
22:47mouth movement.
22:48>> [laughter]
22:49>> So, but now you can see that they made
22:50their initial initial picks, right? So,
22:53we can see that Rebuild for Humans is
22:55one of the main is the winner. I don't
22:56exactly know what that means, but I
22:57guess we'll find out.
22:58>> Mhm.
22:59>> But now they're going to go through and
23:00we'll see them kind of decide who they
23:03want to chat with.
23:05>> And so will they bring in different
23:06people or is it it stays the same panel
23:09at this point?
23:10>> At this point it'll stay these same six
23:12council members.
23:14>> Got you.
23:14>> And we'll just see them sort of make
23:16connections, but of course, if this
23:18localhost wasn't open, this would just
23:20be running right here, you know, like
23:21obviously be doing the exact same thing,
23:22it would create the same deliverables,
23:23same sort of vibe, but um this is just
23:26way more fun, so.
23:27>> Yeah, yeah, so everything's happening
23:28basically in in the sort of AI chat. The
23:31visual is just to to to make it a little
23:34bit more interesting and entertaining to
23:35watch.
23:36>> 100%. Yep. There we go. So, now we can
23:39see that Gary Vee was going after
23:41Hermosi a little bit and when they do
23:43stuff like this, if I scroll down here,
23:45we can actually watch like their opening
23:47positions, we can watch the key points,
23:49we can see the debate. So, Gary Vee
23:50versus Hermosi, he said that he
23:51disagrees with Hermosi saying the
23:54readers are fine and the broken part is
23:55a sponsor offer price on opens. Um I can
23:58see it here. Here are some other things.
24:00Here's my scoring. And so, we can go
24:02ahead and read all of the actual like
24:03debates and discussions that are going
24:04on, but now I just wanted to scroll back
24:06up so we can see, you know, like Paul
24:08Graham also disagreed with Hermosi. And
24:10it looks like Bezos disagreed with uh
24:12Warren. So, once they feel good about
24:14this debate, they'll kind of like reset
24:17once again and then we'll see them
24:18actually do the vote and then maybe
24:19these scores will have changed a little
24:21bit, maybe they won't have changed that
24:22much. But that's what you can see next
24:24to their names.
24:25>> Yeah.
24:25>> So, you can actually see that everyone
24:26voted B except for Warren and Alex, they
24:28both voted C.
24:29>> So, are there are multiple debates
24:31happening at the same time? Cuz you're
24:32seeing all those lines between different
24:34people. So, those like they're happening
24:36happening simultaneously?
24:37>> They pretty much are just happening one
24:39after one at a time, because then the
24:41council can look at what these other two
24:44people discussed.
24:45>> Okay, got you.
24:46>> So, now we're going through round three.
24:48We're going to see them actually like
24:50vote vote vote vote vote. And then we'll
24:52have a little PDF that we can look at,
24:54which is a very very high-level
24:56snapshot. But then of course, if we
24:57wanted to dig deeper, ask more
24:58questions, run it again even,
25:01we have everything saved obviously in
25:03the logs.
25:04>> So, what does the amend mean over their
25:05faces right now?
25:06>> It's basically just when they come
25:08through and they have voted. There's
25:10basically like one proposal, and then
25:13they're either going to agree with that
25:14or they're going to push back. So, it
25:15looks like they've they're all starting
25:17to come to a place of agreement. And now
25:19they're writing the verdict here. But
25:22yeah, if we wanted to really get in here
25:24and look at every single round of debate
25:27or discussion, then it can get pretty
25:29granular in here.
25:30>> Nice. And how brutal are they to each
25:32other? Looks like it's pretty cordial.
25:34>> Yeah, that's why we got to get Elon in
25:35the mix, right?
25:37>> [laughter]
25:37>> There we go. So, we have go with B and
25:40run C beside it, which is interesting.
25:42So, a summary passes along facts, not
25:45Matt's judgment. B is the only option
25:47that tests whether readers want Matt.
25:49Pricing on opens prices on the number
25:51against agency. About 78K a year is at
25:54risk on assumed numbers and click terms
25:57protected at no cash cost. So, we can
25:59actually open up this PDF real quick.
26:02>> Mhm.
26:02>> And that will just like I said, it will
26:04give you the the super high-level
26:06overview. So, this is what that looks
26:07like. It's pretty much the exact same
26:09thing. It'll show you the scoring, the
26:10final call. It'll show you how they got
26:12here and like a finished round scoring
26:15of all of these different options with
26:17their weights. And things to watch out
26:19for, and then it kind of gives you like
26:20an action plan.
26:22So yeah, just a super simple breakdown
26:24of what was actually discussed on the
26:25council.
26:26>> And then, you know, the probably the
26:28easiest way to set something like this
26:30up would be to go into like your Claude
26:32code, point it at the GitHub repo that,
26:35you know, Nate made available for
26:37everybody, and just say, "Install this
26:39for me." Right? Like that's almost the
26:41easiest way to do any of this kind of
26:42stuff these days.
26:43>> Yeah, literally just drop in the URL,
26:45set this up, and then you're you should
26:47be good to go. Yeah.
26:48>> Awesome. We're not even done yet. Nate
26:50is a master at finding killer ways to
26:52use AI, so I wanted to squeeze just one
26:54more idea out of him. So, this time
26:57we're going to tackle social media. In
26:59fact, we're going to put Instagram on
27:00complete autopilot. This next workflow
27:03shows you exactly how Nate is putting
27:05out high-quality content on Instagram
27:07using AI and his practically hands-off
27:10technique. Check this out.
27:11>> So, what we're looking at here is it's a
27:14very humble Instagram account. I'd say
27:16this has been going for about a month.
27:17At this point, 99%
27:20automated, maybe 98. The goal is to get
27:23this 100% automated, and then, you know,
27:24be able to monetize through sponsorships
27:26and things like that, but really just to
27:27have a nice automated Instagram account
27:29that's dropping news. So, before I get
27:31into the workflow, we'll just give a
27:33quick overview of what actually happens
27:35here because there are some things in
27:36here that are pretty cool. So, some of
27:38these, obviously, are just, you know,
27:39grabbing a cool video that may be
27:41trending on X, and just kind of throwing
27:42it in here, right? We're crediting the
27:44original person, and then we're just
27:45kind of ending with like a hey, "Hey,
27:47I'm Moose Fast, so go ahead and follow."
27:49Right? Like nothing too complicated.
27:51Obviously, we had to build a skill
27:53around this so that we have a few
27:54different scenes, and then everything is
27:55very branded. But what else is cool is
27:58if we look at a different one like this
27:59one, for example, the opening carousel
28:01was an AI-generated image. We've got an
28:02image of Dario crying with a bunch of
28:04money, and the idea is that I'm able to
28:07either drop in an X post that I find,
28:09right? I can just literally send Claude
28:10the link, and then it just makes a post.
28:13So, that means that it goes and it it
28:14does research. It'll go to a thread. It
28:16will pull off different reactions. It
28:18will also go like read articles, and it
28:20will either create B-roll, or it will go
28:22to the article. It even found this clip,
28:24right? It found a clip of Dario, threw
28:25it in here with audio. So, it's more
28:27than just making a carousel. I've It's
28:30It's automated like the back-end
28:31research and thought that goes into
28:32making the carousel as well. And I'll
28:34just go through one more thing that this
28:37does. So this one for example did really
28:39well and you'll notice that it has 559
28:41comments and this is obviously using an
28:44automation right like a many many chat
28:45type of automation but I've automated
28:47that too. So we're giving away a
28:48resource guide and we could link people
28:50back to our community if we wanted to. I
28:51haven't done that yet here just giving
28:53away the stuff. But not only are they
28:55commenting for it which helps like boost
28:57engagement and things like that but in
28:59the DMs with these people we're making
29:00them follow first to get the automation
29:02or to get the prompt. So they're
29:04commenting and then it says hey if you
29:06you know drop us a follow we'll send it
29:07over and then we shoot it over. And so
29:09all of that's automated and now like I
29:11said it's you know 99% because every 4
29:14hours it goes through it finds a few
29:15posts and it makes some carousels and
29:16posts them but then I'm also still if I
29:19see something I'm still dropping it in
29:20and I am you know may maybe sometimes
29:22still iterating on the prompts a little
29:23bit. Sometimes I'll see a post that I
29:25didn't like I'll delete it I'll shoot it
29:26back and say hey you know I didn't like
29:28this piece of it do it again blah blah
29:29blah. But yeah I mean this really will
29:31get to a point where it is 100%
29:33automated. You know this previously
29:34would have probably been a a full time
29:36role for someone to manage and run this
29:38Instagram account and once we do get to
29:40monetize it I will 100% have agents that
29:42are managing the sponsorship inbox
29:44automatically taking payments
29:45automatically creating those posts
29:46scheduling them out communications that
29:48is going to be super super cool and you
29:50know I think that just goes to show 30
29:52days almost 2K followers. 2K is not that
29:55impressive but it I think it's a lot
29:56more realistic of what is possible as
29:58long as you're iterating on the product.
30:00Pretty impressive for 30 days honestly.
30:01I mean you know if you told me you've
30:03been working on this for 9 months and
30:04you're at 2K that'd be one story but 30
30:06days in I I think that's that's great
30:08for 30 days. Yeah I know I'm I'm pretty
30:10happy with it especially if you if you
30:11really think about like in total the
30:13past 30 days how many total hours have I
30:15spent on this? I would probably tell you
30:17four to five. Yeah I'm pretty happy with
30:19it. Now what I thought would be we could
30:21do real quick to end this episode or to
30:23end this demo would be I could go grab a
30:26quick X post and just shoot it in and we
30:29can see it run. So I guess you can see
30:30this here's an example I just found I
30:32could become self-aware we've got a
30:34video here and what I can do is bring
30:38this over to Claude. And so what I
30:39always do here is, you know, you can see
30:41that I got out a different one. Now I
30:42just posted and it posted it. But then
30:45what I always do is I keep this on
30:46remote control, and then I keep my
30:48computer awake so that if I'm like at
30:50the gym or on a walk and I see an X post
30:51I can just shoot it off. But also I have
30:53this connected to like a VPS version of
30:55Claude code so I can do it there as
30:56well. But I basically would just say
30:59make a carousel for this, shoot the link
31:01off, and then when I run that, it goes
31:04there, it's going to read the replies,
31:05it's going to, you know, download the
31:06video, slice it up, and it will pretty
31:09much just make the whole research post.
31:11It might decide that there's a good
31:13resource to have people DM for, it might
31:15not. But obviously we can enforce that,
31:16you know, if we wanted that to happen
31:17every time we could enforce things like
31:19that. But yes, it's just pretty cool
31:20like automated pipeline now of creating
31:22these social media posts.
31:24>> And you've got a whole skill built so
31:26when you say make a carousel for this it
31:27just goes and knows what skill to grab.
31:29>> Yep, and that skill has things like our
31:31brand guidelines, it has things like the
31:33way that we talk, it has assets it will
31:35save. So sometimes if it's creating
31:36generated videos or images it will save
31:39them too in case we want to reuse them
31:40later. In this one it's going to it took
31:42the clip already you can see it's going
31:43to transcribe it. And my guess would be
31:46that it's going to cut that up and have
31:48a few like 20-second clips in the
31:49carousel of like important things that
31:51this person said in the clip. But we
31:53will see what it actually does. Here's
31:55another example of like obviously it's
31:56it's taking screenshots and stuff,
31:58right? Because when it does things like
32:00this where it's putting this little
32:02thing behind Dario, it has to like cut
32:03out Dario, put that in there, and then
32:06layer it back. And it has to make sure
32:07that it's not going to cover up his face
32:09or anything like that. And sometimes
32:11it'll just play a video here too. So it
32:13will make sure that it's spacing
32:14everything out right that it actually
32:16makes sense.
32:17>> How long does this typically take to
32:18run? So once you feed it a video and
32:20say, "Hey, make this carousel." What
32:22would you say like the average amount of
32:23time it takes to do it is?
32:25>> I would say it's probably about 5
32:26minutes. Maybe this wasn't the best
32:28example because whenever it's a video it
32:29obviously takes a little bit longer. But
32:31it doesn't take too long. And then you
32:33can also just like I've come into a
32:34thread before and just dropped in like
32:36like like 10 posts that I found that I
32:38liked, and then it will just bang all of
32:39them out. And also the desktop app it
32:42does this cool thing now where it will
32:43suggest to open up another thread and
32:45then it will just like slot in a thread
32:48below this one and they're all linked
32:50and then you can have this one main
32:51thread manage all the other ones. I
32:52don't know if you've played around with
32:53that yet but it's pretty slick when when
32:56you shoot off a ton of work to one
32:57thread it'll be like, "Hey, you know,
32:58these are all kind of separate tasks.
32:59Let me shoot these off in like their own
33:01like parent or child sessions or
33:03something." Yeah, yeah, that's kind of
33:04like the new projects feature in Claude.
33:06>> Yeah.
33:06>> Yes, that that that's what it is, yeah.
33:08>> Super cool.
33:08>> pretty nice. And also that's another
33:10piece that it does is it confirms
33:11everything.
33:12>> Oh, nice.
33:12>> makes sure that everything is real
33:13because sometimes on X there's a lot of
33:15like fake news, right? And then the
33:16other thing it will do is if I'm
33:19manually asking it to make a carousel
33:21when I say post it before it actually
33:23does it it will go make sure everything
33:25is still true because sometimes I've had
33:27it happen where I have it generated a
33:29carousel and then I ask it to post it
33:31and then the old post actually got
33:32deleted or
33:34edited or something else has changed so
33:36it does like one more sweep through
33:37which is nice.
33:38>> Instagram's API is kind of rough, right?
33:41So I'm imagining it's actually opening
33:43the browser in Claude and you're already
33:45logged into Instagram so it posts inside
33:47of the Claude browser that's already
33:50logged in. Is that what's happening or
33:51is it going directly through the API?
33:53>> Actually this one is using Blotato
33:55everything.
33:55>> Not familiar with that one.
33:56>> It's the tool that Sabrina Romanov
33:58created if you're familiar with Sabrina.
34:00>> Yeah, I've heard her name before.
34:01>> Yeah, so it's it's a tool that she
34:02created and it allows you to have
34:05multiple accounts. So like I told you
34:06earlier I've got a few other things that
34:07are auto posting on X and LinkedIn and
34:10Instagram, TikTok, a bunch of different
34:12accounts and Blotato can just manage all
34:14those accounts and then it also she just
34:16pretty much worked in the DM automation
34:17functionality and that's when I was
34:18like, "You know what? I'm going to I'm
34:19going to test this out on you know this
34:20this account."
34:21>> So the DM automation stuff isn't
34:23actually through ManyChat it's through
34:25What did you call it? Blotato or
34:26>> Blotato.
34:27>> Blotato, okay.
34:28>> It's the ManyChat style but yeah, it's
34:30not it's not through ManyChat. It's
34:31using Blotato. So you can see this one
34:33finished up. It took I think it said a
34:35little over 5 minutes like five or six.
34:37And what it does is it gives you this
34:38video so you can actually just like
34:40preview the entire carousel. We've got
34:42the opening slide and it chose to do
34:44this tweet style. So, it played a lot of
34:45the video in just one slide. So, I think
34:48this is a two-slide carousel and it just
34:50ended it with that. So, it was looking
34:52through to see what would be a good
34:53frame for the last picture. But yeah,
34:55then it just basically put this with
34:56these with this caption, "These systems
34:58might acquire some feeling of
34:59self-awareness."
35:00>> Mhm.
35:01>> You can see Oh wait, there are three.
35:03Video tweet 36 seconds and then another
35:05one's 36 seconds. Oh, okay. So, halfway
35:07through the text switches to we're made
35:09of the same carbon matter with our
35:10squishy brains. And then you can see
35:11here it's already recommending to post
35:13it. So, I would just hit tab, hit enter.
35:15And you know what? I might I might as
35:16well just go ahead and post it. But then
35:18it'll do a quick refresh and make sure
35:19everything's still good and then it will
35:21just pop up on my my Instagram and say
35:23that it's live and it also logs
35:24everything on a a Google Sheet. So, I
35:26can go see every single post um and that
35:28way the reason I'm having it log
35:30everything is just so that it makes sure
35:32that it's not duplicating a topic. So,
35:34I've also got that going on, too.
35:36>> You know, what would be cool, and this
35:37is just thinking out loud, is if
35:38periodically it pulled in stats of how
35:41they're performing into the Google Sheet
35:43and then it could go look back at what
35:45stuff performed and almost have like its
35:47own internal reinforcement learning of
35:50how to get better and better because
35:51it's watching the stats of what
35:52performed.
35:53>> Yes, you are you are 100% right and I
35:56forget if that is automated or not, but
35:59here I'll actually just show you real
36:00quick. Okay, so I haven't set that up
36:02yet. But yeah, I've prepped it to do
36:03that so that over here you can see it
36:05will go through and it will start to
36:06come back and understand the stats on
36:07it. I guess it did it weekly for some of
36:09these, but not all of them. But yeah, I
36:11do need to set up that element and then
36:13that'll be really cool to see how it's
36:14able to do like a weekly analytics and
36:16um see if it updates its own skills a
36:18little bit based on the data it's
36:19finding. So, 100% yeah.
36:21>> Yeah, sort of self-improving loop.
36:23That'd be super cool.
36:24All right, that's what we got for you
36:25today. Three workflows that you can go
36:27and implement right now. And just as a
36:29reminder, like the easiest way to go and
36:32implement these is go to the GitHub page
36:34where Nate made all of this stuff
36:36available, copy in the link to that
36:39GitHub page to your Claude code or your
36:41Codex or your Grok Bot or whatever app
36:44you're using, and say go build this out
36:47for me, and it will just
36:49build everything for you. But once
36:51again, I want to thank Nate Hurk so much
36:53for jumping on with me and breaking down
36:56his workflows and going into so much
36:58detail so that I can share them with you
37:01and you have some really cool practical
37:04use cases for AI. This is a series that
37:07I'm going to be doing more of. This is
37:08my show me your workflow series. The
37:10last one with Laval's to do showed off
37:13his God's eye view, and now we got three
37:15more workflows from Nate. If you like
37:17these types of videos, let me know in
37:19the comments and I'll make more of these
37:21show me your workflow videos. And if you
37:23haven't already, maybe consider liking
37:25this video and subscribing to this
37:27channel because that will make sure that
37:28when future versions of this pop up,
37:31they'll show up in your YouTube feed.
37:32And if you're brand new to this channel,
37:34I also try to keep everybody super
37:36looped in with all of the AI news. Every
37:39Friday I make a breakdown of all of the
37:41AI news that happened throughout the
37:42week so that if you don't want to have
37:44to drink from the fire hose all week and
37:46be overwhelmed by all the AI news, well,
37:49I make one video that you can check into
37:51once a week and get fully looped in for
37:53the week. So if that's something that
37:55interests you as well, definitely
37:57subscribe to this channel and I'll make
37:58sure those show up in your YouTube feed.
38:00Thank you so much for hanging out with
38:01us today and nerding out with us. This
38:04was a blast to do and hopefully I'll see
38:06you in the next one. Thanks so much.
38:08Bye-bye.

About this transcript

This page contains the full transcript of Nate Herk Breaks Down His 3 Most Powerful AI Automations by Matt Wolfe, generated from the public captions YouTube serves with the video. The transcript has 9,035 words across 1,263 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.