f*ck it. Here's how I automated my youtube channel in 24 mins (I show everything)

Hey everyone, I'm going to show you three agentic workflows for video. The first is an endto-end AI video editor that actually cuts up my content, mixes and masters it, and then uploads it to YouTube for me. That is what you guys are currently looking at right now, the byproduct of that flow. The second is an AI thumbnail generator. You guys are looking at the thumbnails right now before you clicked on this video. And the third is an AI outlier detector, which is how I came up with the idea for this video. So, I'm going to show you guys how to build these sorts of agentic workflows. I'm going to show you how all the stuff works, and then I'm even going to share them down below in the

description for free. So you guys are looking at the byproduct of this video editor right now, but in a nutshell, this automatically edits my talking head videos by removing silences and then actually enhancing the quality. The way it works is we start by extracting audio from a recorded video like what I'm doing right now in um Open Broadcast Studio, just some software platform. You guys can use whatever you want. It then runs this neural voice activity detection module called Solero VAD. It then removes all the silence gaps between speech. Um, I set it to 0. 5 second threshold, but I don't know, sometimes in videos you need to take a pause or you need to step out change tabs or whatever. This deals with that automatically. If I say these words, I'm not going to say them because it'll actually go through and edit my video. But if I say these two words, it'll actually go and then remove the part of the video that I made a mistake on and then go back to the previous section. And then I do some audio enhancements before applying some color grading. Then finally adding that swivel teaser that you guys saw at the beginning of this video. And then we use what's called hardware acceleration just to process it all much faster than Premiere Pro. So I developed this because I do a lot of videos now. It's a big chunk of my workflow. And to me it's always been really important to economize the video production process however possible. I should note that I know none of the programming involved in this. This is entirely built through cloud code and I've tested this now on several videos, some of which most of you guys here have probably seen. Anyway, most of my subscribers that is um and yeah, the results are pretty incredible. So I just wanted to start with this because I think this is a pretty common use case. A lot of people here want to do videos. Probably sell to people that produce videos and stuff like that. This is definitely something that's really straightforward and simple. The way that I always recommend setting up your agentic workflows is using the directives orchestration execution layer. This is a three-step framework that I've talked about in previous videos, but essentially you store highle instructions about the task in this directives folder. Then you store uh Python scripts and stuff like that in the execution folder. I know nothing about the Python behind this. This is just where Claude code does its thing. And when you do it this way, what you do is separate concerns. Instead of forcing the language model to do a lot of the heavy lifting on its own, language models are pretty probabilistic, which means while they can succeed from time to time, they also make mistakes. You just allow it to call tools that it's already invested all its time and energy into to minimize any sort of um accuracy issues. So, here's an example of my smart video edit directive. I'm actually going to give this to you guys down below. Essentially, it starts by giving a brief description of what it is. So, we automatically edit talking head videos. I then provide a background image when the user says edit X video. Here are the steps to run. Always use the parallel version wherever possible. We have some parallelization in place. Here's what this produces and all this fun stuff. I mean like I didn't actually come up with this. I just ran this so many times and then at the end I said, "Hey, turn what we just did into a directive. Super straightforward. " And then down below we also have a bunch of executions related to this. So, if I just scroll down and then get to smart video edit, sorry, simple video edit. Uh, and my AI should have just cut that out. We see here that we have some comments and then down here the actual script. So, we load in some anthropic API cues, some configuration. Then we have a bunch of different thresholds and stuff like that. It's pretty badass the way that this thing works. I use this in parallel with a bunch of other scripts that you guys could see here. Although obviously I don't actually control or manage any of this stuff on my own. Basically what occurs is I say, "Hey, can you edit my video? " It then looks through my directives folder, finds this directive, and then it just scrolls down here, and then looks to see, okay, what scripts are we calling? Eventually, it finds a bunch of scripts, and then it goes and it calls them all in order. So, I mean, you know, I'm not a programmer by trade. I've done some front end way back in the day, but I'm certainly nowhere near as skilled now as I was back then, and to be honest, back then I wasn't very good. Anyway, what's really cool is this is something that you could pay a software company a fair amount of money. um on either like a monthly subscription or maybe like pay some developers to put together for you that previously probably would have cost me many thousands of dollars, potentially tens of thousands of dollars or more. And now I'm capable of doing this in much less than an afternoon, probably about half an hour back and forth. And I'm going to run you guys through how to do that sort of thing in a second. So how would I actually build this? It's as simple as me literally saying, "Hello, I'd like to build a simple workflow that takes as input a video, identifies the silence gaps, cuts those silence gaps, finds any mistakes I make, cuts those two, then builds a little cool intro animation, which in my case is a video swivel. " Now, I've already built this and it's inside of my directives and executions folder, so obligatory here. I'm just letting it know, hey, there's already a workflow for this, but this is a demo showing building capabilities, so I don't want you to acknowledge or even look at these. Build it from scratch. start by giving me three options we could use. Look at pre-existing solutions and the like and lay them out. If it makes sense, we can proceed. So, I'm just going to head over here and move to plan mode. I'm going to press enter and then I'm just going to allow it to come up with a plan. Now, I usually recommend anytime you're building any sort of workflow or anything like that. The best way forward is not just to come up with one um solution or possibility, but come up with multiple. And then you essentially, the sort of person in the driver's seat just gets to pick the best one. The reason why is because AI can have a habit of just going down little rabbit holes that don't actually go anywhere. It's just important for you as somebody that is sort of guiding the AI to at least be in the driver's seat a little bit. So, it's gone through and performed some web search for pre-existing solutions. I usually recommend doing this as well because odds are somebody will have a blog post talking about how they've done this. And then instead of you rebuilding the wheel from scratch, you can just use theirs. And as you see here, it's delivered me three options. We could use a pure local stack with this software platform and that software platform. Now, in my case, I've actually done this exact same thing before. So, I know what these software platforms are. But let me tell you, most of the time when I ask you to build things like this, no idea what the software platforms are. I just have it run me through it at a very high level. So, then it gives me some, you know, uh um very detailed instructions. I run ffmpeg- af silence detect. Do I actually look at or use any of these? No, I don't. And then I have a hybrid cloud one here. And then down over here, I have a VAD first approach. Now, this is actually what I ended up going with, although I didn't use movie pie. I used some other library, but essentially, um, I tried all of these and I actually took each three of these and then I fed them over to other instances of cloud and I said, "Hey, how's it going? I'd like you to build a video editing workflow that does this. " And you know what I did? I just tested all three simultaneously. And I figured out that this one over here worked the best. Took me, you know, somewhere like 30 minutes to maybe an hour or so in total to do this. I did need to record a video to actually test this on. And there was of course some back and forth, but this is extraordinarily straightforward to build. And as long as you use a framework to help constrain that, I got to be honest, the only thing really limiting your ability to build at this point is just your agency. Can you identify a problem in your day-to-day workflow and then make a decision to solve it? So after we've built said workflow, we obviously need some raw material to test it on. And so that's the next step. I recorded an 8 minute and 51 second clip here called 1_cut_851. So here I just go run the video editing workflow on one_cut_851 MP4. I'm going to set it to bypass permissions that I can go through things. And because I've already stored some highle instructions on telling it what to do and essentially how my um directive orchestration execution framework works inside of this agents. mmd file or claude MD or Gemini MD all just depends on what you're using. Essentially, this knows that when I say, "Hey, I want you to run the video editing workflow," it needs to check directives first and then find the executions listed. And once it's assembled all that information, it actually just goes and runs. So, as you guys can see here, it's actually done this now. So, it's saying, "Hey, found the video. " And then there's a background image running the video editing workflow. The next workflow is essentially a cross niche outlier finder. Now, um, just as somebody on YouTube, maybe you guys aren't in the know, but a big way that people on YouTube choose the titles and the thumbnails to make videos on is they just look to see what are other titles and thumbnails that other people have used in the past that have worked really well. The way we do so is we calculate what's called an outlier score. So, this one over here, um, Elon Musk, a different conversation or whatever. This one scores 6. 81 81 on the outlier score, which tells us that, you know, there's something inherently interesting about this video because it performed at almost seven times the average when compared to other videos on that same channel. Now, in this case, it's probably just because uh he doesn't do a ton of interviews and so that's interesting. So, scrolling down, launching a $und00 million startup from our living room with a little time and then said living room. That's pretty cool. Jack Maw, you know, a big

breakdown of him. This dude over here with thing that says starting. The point that this that I'm making here is this essentially allows me to compile a big list of outlier detectors um of rather high-quality YouTube videos or maybe high-quality packaging for YouTube videos and that I can then repurpose for my own. I'm going to show you guys how to apply that with an AI thumbnail generator in a second. But the way that I built this was super simple as well. Basically, what I did was I used this API called TubeLab, which actually includes the ability to find outliers just out of the box. I then calculated the outlier by dividing the number of views on the video by the average channel itself and applied what's called a recency boost which is just a little algorithm that we developed together Claude Code and I um that multiplies more recent videos by a little bit more. We then added some more cross niche modifiers here 30% if it's a hook that has to do with money, 20% if it time and so on and so forth before fetching the transcript, summarizing with AI, generating some title variants, then outputting into a Google sheet. This is the Google sheet over here as you guys could see and we include both a link to the YouTube video right over here plus we include um obviously the thumbnail, the title, our score and basically everything that I need in order to make a decision. Now, this one was super easy to make as well. All I did was I said I want and I'm actually going to do this live with you using a little um voice transcription tool. I'd like to build a workflow that finds outliers on YouTube. Before we proceed, please look comprehensively across the internet to see if there are any APIs out there that are easily web accessible that we could use that have already done this for us. Look for at least three. And now the idea here is I'm having the agent do the research to find all of the tools. I obviously didn't know about this Tube Lab tool or whatever until before I built this. So very similar and very straightforward. And after a few seconds, it does some follow-up web queries here. One was Appify YouTube Scraper API. The other is TubeLab API access developer pricing. This is the service that I ended up going with. Okay. And then now it's giving me some information down here. TubeLab is up at the very top. Says, "Hey, the pricing is really cheap. You can do this. You can do that. " Ampify is next. And then you can use Knox influencer as well. It's even then giving me some honorable mentions. So what I do is once I have all this information, all I do is I take the winner over here. Okay. And then either in the same um you know claude code instance or maybe in a different one, I just tell it what I wanted to do. Hi, I'd like to build a workflow that uses TubeLab to scrape outliers in niches that are similar to mine but not related. Once done, I'd like you to analyze the transcripts of the videos and then return a summary, all of the fields from TubeLab, whatever they are, and then three alternative titles based off of my own YouTube channel that would apply to the video content that I create. I'm also going to tell it I've already built workflows for this. You'll find them in directives and execution. However, if this is a demo to demonstrate the ability, your capability to build workflows. So, please do not access either of those and I don't even want you to look at them. Once you're done, build a comprehensive plan and put it in front of me. If it makes sense, we'll move forward. Big fan of using u both plan mode and then voice transcription here. Just makes it much faster and easier for me to communicate with models. Like if I had typed all of that, I probably wouldn't have typed it verbatim because obviously um you know, [snorts] you tend to talk a little bit less efficiently, I suppose, than you write. But um you know, you write at like 50 words per minute or something on average, right? And it's like I talk at like 150 to 200 as some avid viewers of my videos will know because they'll say, "Hey Nick, why the hell do you keep talking so fast? I have to play you at 0 75. " Um, the point that I'm trying to make is I basically just get all my thoughts out using some sort of voice transcription service. It's very low friction. I just press and hold a button and then this thing actually does everything else that I need. Cool. So, I really like opening the thinking tab just to see what it's going through um, you know, sort of behind the scenes. You can see that it's asking me some information here like hey you know what's your YouTube channel? So I'm going to give it my YouTube channel. I'm going to say related niches. Um anything to do with business but not necessarily related to AI or AI agents. Next we'll say Google sheet. Volume will say scrape 100 per run. And then filters will say look for videos that were published in the last month or so. Cool. So now I'm giving it all the information that I want it to have. then finds everything I need. It's now writing some workflow steps and it will present this to me. I'm not inside of the thinking tab in a moment. What's really cool is you don't even need to know the platforms themselves anymore. And that was something that I used to have a lot of value in knowing. I mean, I used to spend a bunch of time and energy trying to stay up toate on the best tools and technologies for various things. You know, customers for automation and AI consulting would be like, oh, like what are the best tools? And if I could present them a tool that I don't know, saved them even 10 or 15% over the course of the next year, that made that knowledge so valuable that they were willing to spend several thousand dollars a month with me. Um, now AI agents can just do all of that. Once we're done, we then have some quick questions here. So I will say yes. Let's do claude opus 4. 5 and then create new sheet. I look over here at the big plan and it's more or less exactly what I wanted. So, it's going to build the workflow by quering YouTube outliers, fetching the transcripts, doing LM analysis, and outputting to a Google sheet. I've obviously already built this um several times in a variety of different forms, not just with Tube Lab, with a couple of other tools. So, I know that this is more or less exactly what I want. From here on out, all I do is I say go and then it goes. So, let's go and then it's implementing. So, it's going to start by making a to-do list, building the directive and execution script, and so on and so forth. And at some point it's going to ask me to provide my API keys, in which case I will. But yeah, I mean it literally is that easy. I just ask it to do something. I have a loose idea of what's going on. I have it come up with the tools. I even several approaches. And then I can either choose one myself, execute them all in parallel. I mean like really the world is your oyster here. Once we're done, my workflow is as simple as find me, let's just say 50 outliers, and dump to a Google sheet. Because I use the directive orchestration execution framework, I just give it instructions and then it knows what to do. So, it's now going to use the API to find me what I want. You can see that we've put together that script which is in scrape_cross_niche tublab. py. We've even set a little limit parameter. That's just the code that it wrote. And then once it's done, we have the Google sheet right over here which um I just need to allow access to see the thumbnail. There we go. And now you can see, I mean, because we ran the same search, it's obviously getting some of the same thumbnails and titles, but uh yeah, pretty cool. The final workflow that I put together that I want to show you guys is an AI thumbnail generator. So, I've been using variants of AI thumbnails recently on YouTube, and a lot of them have honestly been pretty poor, but I've recently cracked the code on how to make sure that they look actually like me, not in that weird uncanny valley. And um I got kind of an interesting uh set of realizations to share with you guys. So, the way that most people do AF thumbnails is they'll go into Nano Banana Pro and then they'll put their face and then the face of a thumbnail that they want to replicate and they'll just be like, "Hey, can you replace it with my face? " And this does okay, but um I will say it kind of fails spectacularly a lot of the time. And I've come to realize the reason why is because face direction. If I have a thumbnail that I'm interested in replicating, let's just go back to this sheet and say I want to replicate this one here, The Spirit of Excellence with Pastor Matt Hegy or something like that. Um, and that's sort of like my, you

know, that's what I want to copy. Um, notice how he's looking at me directly head-on. Well, if I were to try and replicate this with, uh, a picture of myself that was not directly head-on, then the AI would have to make some inferences about the way my face looks in different angles. And that's where it gets really crappy. So, I've solved this with an AI thumbnail generator that face swaps existing YouTube thumbnails to feature me using image generation, specifically Nano Banana Pro. The way that it works is it actually analyzes the face direction, aka the yaw and the pitch of like your nose, which is insane. Uh, using this tool called Media Pipe. Then it finds the best matching reference photo of you by uklidian distance in pose space. Do I know what any of this stuff means? No. But you rest assured, I've tested multiple approaches with cloud code and I've determined that this is the highest performing one. Then I send the source thumbnail and two reference photos to Gemini with a face swap prompt. Then generate three variations per run. Um, optionally I run some sort of iterative edit pass for refinements to change the text, the colors in the background. Okay, so I have this one over here and I actually have a link directly to the YouTube video. So I'm just going to copy this over. Then I'm going to go to my cloud code instance and then I'm just going to say, hey, um, create or generate thumbnails like of this using our workflow. So I'm actually just going to include this with no context. I'm going to see how it performs. So first thing it's going to read the directory for recreating thumbnails to understand the process. So you know just instantiated this cloud code instance. It has zero context has no idea what the heck's going on. It's now going to run the thumbnail recreation script with that image URL. So as you can see here this script that we developed actually takes in the URL. And then over here on the left hand side I have like a thumbnails folder which we're going to use to actually find the generated thumbs. Okay. And it's ended up just basically hot swapping my face out one for one with his face. I'm not a fan of sort of just the way that this looks though. I mean, there's just some minor variations here between like my hair, for instance. My hair isn't that bright if you guys could tell. And my face is a little bit big. Maybe neckline's a little bit small. So, I'm just going to have it do some editing here. Also, what's really cool is you can actually run this in parallel. So, if I go back to my thumbnail uh picker here, I'm actually going to grab this one here from Tanner Chedester as well. Sorry for absolutely ripping you buddy, but um let's just do that simultaneously. So run the thumbnail generator on this and it's going to have both of these simultaneously. Okay. And that last one here I said uh change pastor Matt Haggi with Nick change the spirit of excellence with workflows with claude code. And we got something that looks like this. So I mean it is me. Although the hair is still a little bit too brown if I'm being honest. I'm much darker than that. We got ourselves a brown shirt. We removed most of the text here. We called it workflows claude code with Nick Sarif. Um, you know, obviously you can go back and forth as many times as you want, but uh, first demo did a pretty good job. With the second one here, uh, we can see the value in generating multiple. So, I mean, I generated three, right? I always do three. And the first one looks really, really close to me. That's about as close as we can get. Second one kind of 50/50. Looks like an alien neck. Third one, I look younger. A younger, brighter me. Let's just put it that way. So, I'm then just going to say, great. Swap out the text. Swap out um Tanner Chadster with Nick Sarif at Nick. sarif. Swap the profile pick with me. And then I'll say change text from how I make $26,000 a day to how I make agent flows. Okay. And so now we have my how I make agent flows. I didn't change the selling digital products or anything else. Also, I'm definitely not that ripped. Good job, buddy. Um, and then we also have the below thumbnail. And as you guys hopefully can tell, I weave all of these workflows in together. So, like the AI video editor is obviously I'm editing the deliverable, aka what I do here. I record the video and then I publish it. Um, but that's just one tiny piece. In order to come up with the video ideas, I use the outlier detector. And then in order to actually like generate the thumbnails and stuff like that, I just use the thumbnail generator on the specific ones that I generate the or rather on the specific outliers that I find um using said outlier detector. So yeah, kind of cool. Um the reason why I wanted to record this video is not because I necessarily like make content all about video editing, right? To be clear, I just wanted to record this video to show you guys how easy it is to take parts of your day-to-day workflow and then economize the ever loving hell out of them. super straightforward to do this. I mean, in order to make that thumbnail generator, I literally just wrote, "Hi, I'd like to make a thumbnail generator. I want to use Nano Banana Pro. Um, please assist me. " Did the same thing I did with the initial um AI video editor. The thumbnail generator is a lot less complex to be honest. So, all you really need to do nowadays, if I'm being frank, is just like list all of the stuff that you actually do on a day-to-day basis, and then just get like the SOP, get like the deliverables, and just feed that into Cloud Code or a similar agent and just say, "Hey, I want to make a workflow that does this. Can you help me? What are the top five best approaches? software platforms to do this? " Once you're done, give me a plan. Then you just like pick a plan or you just run them all in parallel, test, and then find out um any costs that you will incur during this process. And you probably will incur costs like I had to incur some API costs probably about $25 in total to build that workflow editor. Um any cost that you guys incur in the process will be marginal relative to the amount of time, energy and money you are spending probably to do that for even one day. So it has never been easier to like take that potential of AI than actually apply it to your day-to-day to make a massive change. And I highly recommend you guys get on it as soon as humanly possible. Okay. So I'm going to include a link down below um with all of these workflows, at least the directives and the executions. Feel free to copy and paste these. I don't care. Um, you know, they are free. Hopefully you guys see the value of or rather software as a moat is nowhere near as valuable as it used to be just a few years ago because you can now just like build anything like this with some bullet points. So, it's not like I'm really losing out too much. You don't need to sign up. give me your email address. Everything's uh totally free and included. If you guys could do me a solid, definitely check out Maker School. It's my 0 to u one 90-day accountability roadmap for getting your first paying customer for a service like this. actually hold your hand, walk you through with daily Loom videos, daily Q& A, and you actually get full access to me. I'm in the community every single day. One of the most active people on school. Um, yeah, definitely check that out. If you guys don't achieve what you want to achieve in 90 days, you guys get your money back, no questions asked. So, that's fun. And then if you guys could do me a big solid and really like what I actually wanted to ask you guys here was if you guys have any workflows or any things that you currently do in your day-to-day and you're curious about whether or not they can be automated and for whatever reason you don't have the wherewithal to try this on your own, just drop a comment down below with it. Like I'm more than happy to like walk you guys through how to automate that process. Um my main goal with this is just to show people that it is all entirely possible. Whatever your day-to-day is, you can probably automate 90% of it. You know, I get a lot of people asking me questions like, "Hey, is AI capable of automating my job yet? " And the reality is it's probably not capable of automating 100% of one person's job. No, but it is capable of automating 90% of 100,000 people's jobs. And that's very much where we are at right now with current levels of capability. So yeah, I'd be really interested in uh you know, working through some challenges. If you guys have some sort of challenge, just let me know down below and I'm more than happy to give it a look. Thank you guys as always for watching and I'll catch all you on the next video. Juice.