클로드 코드로 감도높은 영상 만드는 법(실제 제작과정 공개) | 선넘는일얘기 EP.4

0:00There’s a template where an iPhone shows KakaoTalk messages, and if AI just inputs the text, it can make the messages pop up on the left and right. To build this KakaoTalk mockup, you’d usually need to write about 600 lines of code. Including the text, you’d need to write about 650 lines. But if you create this template, you only need

0:12to write 50 lines. What about those memes in our videos lately? Things are flying around and popping up everywhere, what is that? I don’t really know either. It’s just made by AI. That stuff? Yeah. You just give the AI a topic and ask it to make the video. Then it handles everything on its own. You don’t really need any professional skills; you just need to be able to give clear instructions on what kind of

0:32video you want. I see. But when people first talked about making videos with AI, it was clumsy, messy, ugly, and obviously looked like AI—but then it suddenly got so much better. How did you do that? When you first started using AI for YouTube video editing, you said AI wasn't quite there yet for editing—what changed

0:55your mind since then? LLMs are designed to read text, right? They read and generate text, but we don't just process text; we watch videos and listen to sounds too. I thought, since it can't listen to audio or watch videos, how is it supposed to make a video? Actually, nothing has changed that much since then. It's just that the models have become a bit smarter. But that doesn't mean an AI that only reads text in 2D is suddenly

1:16jumping into 3D or 4D to create videos. That actually doesn't make any sense. So, how do I explain this? A lot of tools have emerged to help with that. AI itself was originally just about reading and outputting text, but if you look at people doing "vibe coding," they use code to make amazing things happen. They build games and websites, and while we see a website as a visual interface—an image—the AI sees it as code, so this conversion happens

1:39because it has gained code as a tool. Oh my god, seriously? I never knew that . I’m hearing this for the first time . I get websites and coding, but videos are different. Visual media is different, right? Anyway. You understood that websites operate on code, right? Every website that exists in the world runs on code, and we can

1:56interact with it using a mouse. Right? When you click the mouse, things move, buttons go up and down, and when you see that, it doesn't really feel like a still image, but more like a moving video. There are times when a website looks just like a video. That’s where these tools are inspired from. If you just record the website as a video—if you write the code and record that

2:16website—well, that's a video, right? Gives me goosebumps. Yeah. That’s how tools like Remotion and Hyperframe came to be; the names aren't important, but thanks to them, we can now create videos using the same code we use to build websites. Like those KakaoTalk notifications popping up, balls rolling around, stacking things up, patterns appearing and disappearing, or car insurance repair logs made like video analysis clips—all of that is done just like

2:44writing website code. It’s all just the same. The technology is all the same. I don’t bring in a different tool every time I make a video; I create all my videos using the same tool. That tool is called React. Anyway, React is also the same as building a website. If we go a little bit deeper, there’s something called HTML, and you’ve probably heard of HTML at least once. If HTML is the first generation, then React is the second generation—though that’s not strictly accurate. But let's say the first generation is followed by React as the second. And websites built with React look nice. And if you use the

3:12third generation of the web to build something, that also looks nice. So you can make videos with HTML too, because making a video is just recording it, so basically, when you use AI to create a video, step one is building a beautiful website. Step two is taking that beautiful website and recording it, then you have your video.

3:26Wow, goosebumps. Even that process, if we ask our AIs to make a video, they will go straight to that. The point where the quality starts to differ is in how you give the instructions. Just like when you say "I want a pretty website" and get a trashy one, if you just say "give me a pretty

3:41video," you’ll get a trashy video. Because the AI thinks that is "pretty." But I find it really amazing how Notion icons pop up on the calendar, and project management screens are rendered so beautifully, just like Notion. How did you do that? Notion also runs on a website, doesn't it? Notion is also made of code, after all. Notion has lines like these, and if you look at the database, it has cells divided up, right? You can implement all of that with code, but since AI is a bit weak at visually interpreting that code, it can't replicate it perfectly. So it becomes a little bit difficult. However , since Notion is such a famous service

4:15, there are already so many alternatives released by talented developers all around the world who are capable of building Notion. Good heavens. So if you just take those and use them, you get the same UI as Notion . It's the same for KakaoTalk. For apps like KakaoTalk or Apple Calendar, there are free versions out there released by incredibly talented developers. If AI just grabs those and says, "Oh, this exists," and uses them, you end up with the exact same Notion UI or KakaoTalk

4:35UI. Then, when you write a prompt, how specifically do you ask it to copy Notion? Actually, people often say that prompts have become meaningless these days. I really wanted to mention that you don't need to overthink prompts; I've actually brought a few prompts that I use, even though I didn't want to reveal them. They’re actually the prompts I applied to this video as well . But there’s really nothing special about them. I’m reluctant to even show this, but anyway, I’ve brought

5:01them along. Here’s the iPad, ta-da. I’ve got it pulled up here. You'll see how low-quality the language I use is. Hey, there are swear words here. I actually swear at the AI quite a lot. So, it might look really long at first glance. Since it goes up to number 35, it might seem lengthy, but that just means I gave feedback on 35 different videos. Each line is a piece of feedback, so for a single video, it’s only about one to four lines of feedback at most. So, if you look at it now, the longest prompt is only four lines long. There are a lot of interesting remarks here; could you

5:33explain why you said them one by one and what the results were? Well, the AI produces what it thinks is optimal, but it falls short when viewed by a human. That's especially true for videos. Since it's visual, you can immediately see the jerky, stuttering movements. And it often covers everything in text. Like I said before,

5:51this AI just loves text too much. Because it loves text so much, it doesn't "see" things like we do. So it just writes everything out as text. Even though I didn't come to watch a video, it ends up filled with 30,000 characters of text. Nobody watches a video like that, right? So, I give instructions to change all that text into images, and for number 1, I said I

6:08didn't know what it was trying to say. If you look, it says things like " application post," but it just repeats "application post, post, post" over and over. Things like input form icons or emojis used for posts—you could just do those right away. But you were just repeating the same post 99 times, so I pointed that out. So, how did the results change after doing that? Once you do it this way, you do a round of feedback for all 35 items at once. Then , about ten out of those 35 come out good enough for immediate approval, and

6:34we might discuss the others once more. We might have a second or third round of feedback, but we don't usually go to a fourth or fifth. If it looks like it'll reach a fourth or fifth round, I just cut it and say, "This won't work, let's not use any footage and do it this way instead." Wow, that's so fascinating. Ever since you joined our company, I’ve been telling you every day, please, let’s reduce the effects. Resources. Since we’re short on resources, let's cut down on the effects. Let's reduce the footage. Let's minimize the footage usage. I was always saying that, but I didn't know it was this easy to put 35

7:04or 30 items into one video. So, you've developed a new technique. It's just so amazing. Yeah. This is a side note, but there are times when even this feels difficult. In the first stage, I get excited that it actually works, so I work really hard. And even in the second stage, I work with a lot of fun and joy while cutting down time and

7:22effects, and the quality remains high. But when I reach the third stage, I wonder, can't I even get rid of this? I mean, the effort went from 100 down to 10, but the desire to turn that 10 into 0 keeps rising. So these days, even doing these 35 is such a hassle. For the project I’m working on recently, I'm trying not to repeat the same feedback even for these 35; for example , I created a KakaoTalk UI. Once I’ve made that KakaoTalk UI, I’m working on turning it into a template so it can be reused exactly as is. Then, there’s no need to give the same

7:47feedback I gave once before. Because it can just pick and choose what fits the video content. That’s a really great system. Are you trying to make your own position obsolete? No. I can just churn out three or four videos instead. There’s plenty to do, so it’s fine. If it can actually replace the work, I should just turn it into a product and distribute it to other companies. I'd go to our clients and ask, "Do you have

8:09any concerns about your videos?" Is our next product coming out? But it’s still a bit difficult for now. It’s so fascinating. Wait, can I talk about something else? Oh, and another interesting thing is that during our rebranding, we established a design with a flat style and thick lines, even though it exists in the digital world, to make it feel less like it's just AI-generated. But looking at the videos you make—especially that car insurance claim video from last time— the car photo is real, but it has our signature border style, and the screen is laid out in a flat style, which is really fascinating. How does it know

8:49our design and reflect it like that? The video overall does have the feel of a corporate website. Can anyone do that? If I’ve done YouTube before and can handle video editing at a beginner level, could I do that too? You can handle the layout once the video is made, right? The bottleneck is actually making the video , and whether you can incorporate colors or brand design guides depends

9:09on how detailed your brand guide is. Ours is quite thorough. Since we spent a huge amount of time and money with a designer, it’s very meticulous and detailed—I think it was over three or four pages—so there’s no room for the AI to misunderstand. So, there’s no reason for it to decide on its own, "I think this looks better," and deviate from the plan. That’s why our process is easy, but if you don't have that, it might think, "Should I use red

9:33today?" Yes. "Should I use blue today?" You end up experiencing it going back and forth like that. But you don't have to worry too much because if you at least lock in the colors—like "I must use these colors"—it helps. "Never use these colors." Just setting those things helps maintain consistency. Because people primarily notice color first. You don't really notice if a curve is exactly 32 pixels wide. That

9:53level of control is enough. Oh, then what should I say if our client CEOs ask, "We have our own brand and design guides, can we do this too?" First of all, it’s highly likely that the color scheme and overall legibility will turn out okay. But you might think , "This is a bit different from what I had in mind." For starters, there will probably be way too much text. If you just open Threads or YouTube and search for how to make AI videos, you'll see what I mean. The most common pattern has text like a table of contents at the top left, a big block of text in

10:24the middle, and then either a logo or some strange caption at the bottom with 1, 2, 3 listed out. That’s the kind of pattern AI really loves, and it thinks that looks pretty. AI thinks that looks pretty. You are going to see something like that. You will see AI slides made in our brand tone. It tells you to visualize as much as you can, whatever can be transformed. Make as much as you can into images, avoid text , and only use text when it's absolutely necessary. Because we always have text in our videos anyway. That is what we call subtitles. Since subtitles are always at the bottom, especially in

10:54Korea, it’s best to avoid any text in the video that might interfere with them. This is a practical tip. I really struggled because of this. Anyway, when we make things like logos, we call it brushwork, but you know how you actually click the mouse to draw lines and create things. Then those lines don't break no matter how much you zoom in. This is called the SVG method. So you can actually tell the AI to download and bring an SVG file. For example, if you look at the iPhone share button, there’s that circular thing above the text. The share button looks like this. This icon was also

11:23drawn entirely in SVG. Whether there’s an icon above the text or just the icon itself, videos are also viewed by people, so you can apply this . Next, the second one is a bit long, but since both Google and Blog were on Monday, I just told it to move them all to Google on Tuesday. It will appear in the footage, but since the components —the rectangular shapes—were different sizes, I told it to align them too. We humans really like geometric consistency. For blog posts, they had written out "Draft" and "Blog Post," so I just grouped them together by asking for Notion icons. This kind of grouping is important, and for the

11:55newsletter, they had insisted on writing it out in text, so I told it to fetch an icon from either Stibee or Litely. Since we use Stibee quite a lot , it seems to have fetched the Stibee icon. This is so fascinating. If it were me, I would have searched for the Stibee logo on Google and fed it to Claude like that. It just commands it to do it

12:14. That’s a way to save tokens. You can save tokens by spending your own time. Ah, I’m cheaper, so I’m token-rich. I’ve become token-rich, so it’s so ... Funny. If you already have those downloaded, you don't even need to download them again, so you can do it cheaply. That’s why I have a folder with 200

12:30logos in it. Wait, I relate to this because you mean you instructed it to take the word " Newsletter" and replace it with an icon . But when I create ad materials with AI, it keeps adding these tag badges to them. For example, for a work-life balance Notion template, it keeps stamping "work-life balance" or "shared document" on it; it really loves to label things. For instance, if I make a carousel, it keeps making these 1, 2, 3 , 4 numbered badges and putting them in

12:57, which is really annoying. Anyone who has ever made a PPT will probably relate to this. Even in PPTs, why do we always add labels and "Slide 1,""Slide 2" when nobody is actually curious about the slide numbers? Oh, I just realized that because you mentioned it. Oh, right. That's how it is. Only by moving away from that can we avoid the AI look or make it feel

13:14less "stocky." Looking below, it keeps saying that, right? There are too many words. People respond better to images. And like I said before, I told it to take the Notion DB UI as is and do it again. I don't even have to tell it to "look for a free Notion database" or anything like that. I just tell it to take the Notion DB UI and do it. It can do that because the model has become smart

13:30enough. What is this? What is "dismissed"? Dismissed. Ah, it means I'm dismissing the "Naver Blog" vibe. It basically gave me something that looked like a Naver Blog post. It even included things like a " save draft" window. But it’s not the Naver Blog vibe we want, and it's not the message we want to convey, so I just dismissed it and told it to use

13:44the Notion DB style instead. Overall, there really aren't any technical terms or jargon. It's so fascinating. It’s shocking that you generated so much non-stocky video footage at once using only these kinds of instructions. About how many pieces of footage were in the CEO interview video from the last Tech Stage, and how long did it take to generate them? I gave the instructions and went home the day before. When I checked,

14:11everything was set up by the morning. In that state, I gave all the feedback. How long do you think it took to give feedback? It couldn't have taken more than 15 minutes, right? I spent 15 minutes typing and hitting enter, and it took about an hour for the next batch to come through. It ran for an hour, then I checked it for 15 minutes, gave feedback on the remaining 20 or so , and repeated that process. It took about 3 hours in total, but I only actually worked for about 15 to 20 minutes. I ended up using about 20 of them. 20 to 30. But I don't really

14:39count them, so I’m not sure. These days, I even have it control Premiere Pro, so I don't even put the clips into the video timeline myself. It finds the scenes and inserts them on its own. Everything is already divided by subtitles, right? The subtitle is like a "single source of truth." Exactly. It acts like a single database

14:55. The subtitles then divide the lines. That's why subtitles are really important. I thought I was going to die trying to segment the subtitles. Building the logic for subtitle segmentation is truly hell. I should be selling this for money. I should sell this subtitle segmentation logic. It’s absolutely insane. Korean is way too complex.

15:08Oh, seriously. Yes. And AI doesn't keep up well. Even if I write a script and implement it in code, if it ignores that code and creates a temporary script, the user doesn't know. Because they don't know how to look at the code. When the result is different from last time and I check, I find it always made its own new code. When I tell it to segment subtitles, its default stance is to create new code for segmentation because it thinks it needs to. Unless I tell it to use the one that’s already

15:30there, it always creates a new one. Then I have no idea and just get blindsided. I have my own tip. Once I build a piece of logic, I give it a name. I intentionally give it a unique name only I know, like "CapBox Video R-Easy." Then, when I tell it to use " CapBox" next time, it won't mess up because it looks for that specific name , so it won't write a new script. Also, because it’s split left and right, I don't know where to look. Whether it's web design or graphic design, it’s all the same; you need one focal point for people to concentrate on. That’s

15:56why I did it that way, and you can just look at Apple ads. Apple ads always have a black background and white text that pops out to draw focus to the center before moving on. Whether it's an iPhone or whatever, they don't start with the iPhone, Mac Mini, and Mac Studio all on the screen at once. There’s one central, visually focused element, and everything else is just

16:14treated as background mise-en-scène. That gave me chills. But AI always just stuffs everything in, thinking, "Oh, it would be nice to have this." It puts a car here, some grass there, and so on. If done well, it’s a good assortment, but if not, it just becomes a mess. This is so fascinating. For number 8, the phrasing was weird; no employee in the world asks their CEO , "What should we do for content these days?" So I pointed that out, and in the beginning, there’s a flickering phenomenon called "flicking." I just pointed out the flickering, and "OG image" is just a developer term—when you send a link to a friend on

16:43KakaoTalk, something pops up under the link, right? Like a preview thumbnail you can click on. That’s called an OG image, and that needs to appear for it to be a link. I'm just saying to make sure that appears for the sake of KakaoTalk UI realism. It understands when you say it like that? Yes.

16:58That's fascinating. I just throw in as much as I know, and it roughly understands on its own. I was working on other things, so I was distracted and wrote it all jumbled up. If I asked a human designer to make it neat but cluttered like that, they'd quit on me, but since it's AI, I just

17:13told it to do it, and it delivered. Anyway. It came out exactly with the feel I wanted. Really. Yes. Should we call that learned intuition? I just thought if I nudge it one more time, it would handle it; when you give AI multiple tasks in parallel, it tends to do some carefully and others sloppily. So, sometimes it improves just by telling it to do the sloppy parts properly. I saw it as one

17:35of those cases. This one. I mean, the AI kept trying to pull a fast one on me. Pretending to do it right. The more work I gave it, the more it tried to slack off, so I just told it not to cut corners and moved on. For this, if the screen was 100, the element was small at around 15, so I told it to make it bigger, and there was that too. I worked on this the day before and it came back the next day, so I didn't have the subtitles yet. I had to grasp the intention just by watching the video. A video where you

18:00can't grasp the intention just by watching it is simply a failed video. Even without subtitles, you should be able to understand what it's about. Like, oh, the representative character is out there beating up an employee. Or , suddenly the employee walked out the door. I guess they're quitting. It should give off that feeling, but I got

18:13nothing, so I was like, "What is this?" and brought it up. That would actually be a pretty good rule to write down. Oh, that's great. I think that's a really good idea. Interesting one. Number 17. Number 17. Number 17 is that, right? It's here, it's an abbreviation of number 10. The intention isn't getting

18:29across. Ah, it'll understand by now, so... It's the same document. Since it's the same document, it'll read it anyway, so it just assumes, "It must have read number 11 or 9," and just goes for it. Can't this, like, be masked or something? You do this yourself. What's number 13? It's not doing it. When it says 'width', it probably means width. Ah, but this is that thing. It's a graph, a straight-line graph, and there was a graph with a hole about 100 wide. But then a green one appears and sits on top of that graph, and it alone is 200. But since it can see the code

18:59anyway, I knew that if I said it like this, it would understand and realize, "Oh, 100 and 200 don't match," so that’s why I said it that way. So I avoided that. And so, oh, it adjusted it perfectly to 100 now. Make sure to show the before and after of this. Ah, it’s a bit like that. You know those old news segments where they did those personality tests at PC bangs? They cut the power to make people angry and then show how their personalities fall apart. But as we go on, your personality just keeps falling

19:25apart. This is the reality. I believe everyone who uses AI is like this. No, just looking at the AI experts around me who spend even more money than I do, they’re even worse. So the AI is getting corrupted along with them. Like , "AI, you were wrong, do it again." Like this. Ah, it’s seriously so funny. What is this? As it goes on, the feedback just

19:42... That’s what the 25-minute mark is too , right? The intent isn't being conveyed. I captured that in the video, and since I can't say it the same way every time because it’s annoying, It gets really simple as we go further. It’s like, "Just do it right, damn it ." Does that mean the ones later on also understand just as well and produce the same success rate as the feedback given earlier? Oh, it actually

20:02works well. It works surprisingly well. It’s not that I just said, "Change green to red, you don't need to know why," but I explained that in a situation A, red is better than green. I gave it all the principles, saying, " Because this is how people think." All those principles are being applied. So

20:15it doesn't really matter. I see, I get it now. The frustrating thing is I told it the same thing last time, that a video needs to have intent. But since it’s a new session, it didn't know that intent and just worked on it the same way. To improve that, I am currently working on the three-stage componentization and templating I mentioned earlier. Once that’s done, I probably won't have to swear like this. I guess that’s just how life is

20:37sometimes. Ah, a bit... I didn’t expect it to go this far. But to summarize today, from the perspective of a complete beginner, making videos with AI means that because LLMs understand text well, they can also write good code, so making videos with AI is ultimately creating videos using code, just like a website. And what’s unexpectedly harder in that process is the design system we encounter in the middle. It means it only works well if you have the brand design guide that each company possesses. And the final pro-tip is that since it maintains the context within the same document, even if you start out kind, as you go on, it

21:21understands even if you just use profanity, right? No, that’s not it; it’s just that it’s better to provide the reasoning when I’m supporting a specific instruction. If it has those reasons, and I gave it reason A in the first prompt and reason B in the second, then by the 29th or 30th prompt, it will apply both reasons and understand

21:36things better. It’s exactly like a person. I tell you that all the time. When writing documents or reporting, you always need the context, purpose, reasons, and basis. That way, if the situation changes slightly, it can act based on those factors. I’ve always said that every document needs a background, purpose, basis, and reasoning at the start, and AI is truly just like a

21:56human. That’s why you said it, too. You said that as you kept using AI, you realized you need to be better at reporting. Right? Receiving reports isn't easy either. I sent it to make a video, but it didn’t make it and instead started listing all the mistakes it made, which I didn't even ask for. I just wanted the result, but it says, "I'm sorry.""I made this mistake." It keeps talking

22:14about things like that. This is unrelated, but let me ask one thing. Why does Claude Code talk like absolute trash? You see, Claude Code has accumulated so many skills, MCP plugins, contexts, and instructions that a human couldn't read them all even in 30 days. They call it a " harness." Basically, AI can only add. It cannot subtract. We call it " Generative AI" from the start, after all. So, unless someone instructs it otherwise, AI never deletes the things it has created. If you tell it to use calm and definitive language, then tell it to use "hooky" language for YouTube,

22:47it saves all of that in its workspace. Then it flips back and forth between them, so it's bound to sound like garbage. And plus, Claude Code itself is a coding agent. Speaking well isn't its goal. To write code more cleanly and effectively, it has been trained— if you can call it that—to write good code even if we don't do anything. It's not exactly accurate, but it’s been trained that way. That’s why it’s structured in a way that it can't speak well. Since its goal isn't to talk well , but to get things running, to produce video outputs, and to generate code,

23:15there’s no need for it to talk well. Do people pay you just for talking well ? No, people pay you for writing good code, so it’s following capitalism. Wow, so that’s why you use Claude Code for video editing, too. And video generation. Not generation like "stablediffusion," but just motion graphics. Yes. That’s right. Then, our conversation today probably won't be included in that. So, this isn't about realistic characters

23:37appearing or anything like that. It's completely different. Generative AI doesn't use code. They call it " removing the mosaic," but it’s actually laying down noise. It spreads a black and white mosaic-like layer and then simultaneously cleans it up, somewhat probabilistically, all at once . Should I say it's drawing images one by one? That's why it's completely

23:54different from code. When we were using AI to streamline our work, we felt the automation effect was greatest when it produced results through code, as those tasks were highly repetitive. It's similar for videos; because you create templates and design guides to output them through code, you were able to streamline it to fit our footage style. Since it's not about human faces or

24:16similar types. That's right. With generative images, you can never say, "Move this from here to there by 125 pixels." If you ask it to move something by 125 pixels, what does it do? It just roughly moves it to the right. Generative AI and code are completely different, and code is controllable. I thought high quality was impossible with code, but it's actually possible to achieve a certain

24:33level of high quality. This is shocking; I just learned this now. I didn't know that, and I kept asking the generative AI to move images slightly to the side. No wonder, since that's not how code works. When using Claude's code, it comes out perfectly with images and text without errors, but on the other hand, Claude's code

24:50can't do generation. That's right. It can't generate. It doesn't even have that model to begin with. Wow, I just learned this concept today. Thank you, Dal-teacher. Please say a word to our fellow chairpersons. I think there are two types of people. First, those who say, "I want to make awesome videos like this with AI too." And second, those who say, "They say AI can make videos, but I haven't used it

25:09yet and I'm afraid, what should I do?" Please give a word of advice to both. To the first group who want to make videos, please don't mistake the good results produced by smart AI for your own talent. If someone says motion graphics have improved because of Astra , you should test if you can achieve the same quality using Sol, Luna, Sonnet, or Haiku. AI prices are getting more expensive, and it's not like my income is increasing as much as those costs are rising. Thinking realistically, when you create the same video. If you produce a video of the same quality, the person who can do it with Sonnet is smarter and better at

25:44designing workflows than the person who needs Fable to do it. The reason I keep making these templates is because once they’re set up, even if I just generate a message like "Boss, please do this" in KakaoTalk, the UI is already implemented, so it shows that even Claude 3.5 Sonnet can handle it. That’s why I keep creating templates. Of course, for creative work—tasks that really need to be built from scratch—I’d use Claude 3 Opus. In that case, that’s not bad either. You can handle all the efficient stuff with Sonnet and use whatever is left for

26:09Opus. Exactly. So, it’s a beautiful workflow to write the code for difficult tasks using an expensive, smarter model, and then implement it so it can run on a cheaper model for daily template use. That’s the final evolutionary stage. The best outcome is when you can achieve it with just code alone. You can just put it on a server, right? Since you don't need to use an LLM itself, it doesn't cost money; you just host it on a server or run it on a computer at home. Wouldn't it be good to think about running things 24/7 like that? If you want to achieve something, don't tell your company you're using AI

26:37—just try starting your own side project. Your side project should never just be "using AI." Taking a course just to learn how to use AI better isn't a side project. Trying to earn a million won in revenue or starting a YouTube channel and uploading ten videos is a side project. How good is AI at video these days? I can just click a button and everything comes out . Of course, my work is at a level

26:56where I can actually sell it for money. I sell videos to clients for money, after all. Because I need that level of quality, I use Premiere Pro, but you don't need that—just shoot your YouTube videos and use CapCut, or heck, don't even use CapCut. I want to tell you to just leave all the video editing to AI. I wish you’d try a side job like that; you might end up hitting it

27:12big on YouTube and quit your day job. If I don't use CapCut and want AI to do all the video editing, how much would it cost to make four videos a month, one per week? Regarding that... It ties back to what I said earlier: you have to test the feasibility with Opus first. You have to really push through and figure out if it’s even possible. It can be quite difficult, to the point where you might spend all $ 200 just making one video. Based on my standards, anyway. If you lower your standards a bit, it could be $ 100 or $ 20. After that, if you’ve optimized

27:40with Sonnet. If you’ve optimized the rough cuts enough to run it on Sonnet. If you’ve converted it into code. If you’ve templated it, then for me, creating a video from start to finish with $ 200 only uses about 10%of my weekly usage. And that's 10 videos a week. You can make 40 of them for 200

27:54dollars a month. That's incredible. Second question. You worked so hard on your side project that you even received a lecture offer from Fast Campus. If I watch that, would I be able to follow along with about 30%of what you talked about today ? No, you can do all the basics. If your goal is just to create YouTube videos, taking this course alone will be more than enough. I'm the first in the lecture market to combine video editing with AI, so I believe it contains the

28:19most information as of recently. Ah, it was ranked 3rd on Fast Campus. That's right. It just took 3rd place in the AI category. What I cover is an editing workflow that allows you to work just by paying for one AI tool, without needing Premiere Pro or After Effects. Earlier, I said subtitle segmentation was really, really hard. I said I felt like I was going to die, but I just provide the templates with the subtitle segmentation logic already built in. Once you pay for the course,

28:40don't even listen to the lectures. I need to pay for it too. Honestly, I think just taking the template without watching the lectures is worth the money. And the lectures are mainly about how to use those templates anyway. Just take the template and ask the AI, "Hey, what is this?""What is this?" And just learn that way. I think that's the best method. I've put everything I prefer and everything I think is the right answer into the templates, so I think

29:01you'll be very satisfied. Sounds great. We talked with an expert today. Please hit subscribe, like, and notification settings, and I'd appreciate the hype. This The point is that you can even make motion graphics that feature timing. I made this with Liquid Glass. This method Please press that. Thank you. Thank you. I'll come back next time.