Как Делать Вирусные Мультики-Шортсы через Claude (Полный Процесс) — Transcript
Full transcript
- 0:00Take a look at these two cartoons. The
- 0:02one on the left got over 30 million
- 0:04views. And it's not just a one-off viral
- 0:07hit.
- 0:07Following that on the same channel: 21
- 0:09million, 16, 15, and 14 million.
- 0:12The one on the right is mine, and it
- 0:14wasn't made by a video AI, but by Claude
- 0:16using 3D code.
- 0:17Frame by frame. Today I'll show the
- 0:19whole process of how Claude studied this
- 0:21style frame by frame, how to build your
- 0:23own cartoon, and what went wrong.
- 0:25And at the end, I'll put these two
- 0:26cartoons side-by-side, and we'll see if
- 0:28it lived up to the original or if it's
- 0:30just a cheap copy. Right away, I’ll say
- 0:31I don’t have any courses. Don't message
- 0:33me privately about training. I’m just
- 0:35showing how I do it myself. Want to
- 0:36follow along? Subscribe to my YouTube
- 0:38channel. And I've posted all the prompts
- 0:40from this video and the ready-made skill
- 0:42for Claude in my Telegram channel. It's
- 0:44free, the link is in the description. If
- 0:46you haven't seen these Shorts yet, this
- 0:48is them: the characters are literally
- 0:50flat, like stickers. Round heads, stick
- 0:53limbs, while the world around them is
- 0:55real 3D.
- 0:56And the camera seems to fly between the
- 0:58objects. That combination is exactly
- 1:00what gets millions of views. So, how do
- 1:02we teach Claude to make videos like this
- 1:04? First, we need samples. And not just
- 1:07one, but many.
- 1:08First thing, I open code, create an
- 1:10empty folder for the project, give it
- 1:12the link to this YouTube channel, and
- 1:14input this task.
- 1:16Download the latest 20 Shorts in this
- 1:18style, the ones that look the best. And
- 1:21here's the most important part: if you
- 1:23just tell it to analyze the video, it
- 1:25won't watch every frame; it will grab a
- 1:27few frames from this clip and maybe from
- 1:29another.
- 1:31But a normal video has 30 frames per
- 1:33second, so I write to it directly: "
- 1:35Watch all 20 clips at 30 frames per
- 1:37second and check literally everything:
- 1:39lighting, contrast, shadows, motion,
- 1:41effects, and cuts."
- 1:43This prompt will also be available in
- 1:44its entirety in my Telegram. And now we
- 1:46have the 20 videos downloaded.
- 1:49And it immediately assigned 10
- 1:50assistants to review these very frame
- 1:52sheets.
- 1:54Why is every single frame so important
- 1:56to us?
- 1:56You've surely seen these cartoons in
- 1:58your feed, and most of them actually
- 2:00look quite cheap.
- 2:02It all comes down to the simplest
- 2:04details: a barely noticeable movement, a
- 2:06slight camera zoom, a touch of shadow,
- 2:08color, and so on.
- 2:10These are what make the image feel alive
- 2:12, and you can only catch them by looking
- 2:13at every single one of the 30 frames.
- 2:16In total, 20 Shorts, almost 19 minutes
- 2:18of video, over 34,000 frames, and the
- 2:20end result is a ready-made rulebook for
- 2:23Claude in the style we copied.
- 2:27The secret to these Shorts breaks down
- 2:28into three layers: how the text is
- 2:30written, how the visuals are assembled,
- 2:31and how Claude measured it all.
- 2:34Let's start with the most obvious text.
- 2:36Claude extracted the audio from all
- 2:38twenty videos for us and transcribed
- 2:39them. And more than half, 12 out of 20,
- 2:42start like this. A man was buried alive
- 2:45with a bell on his finger. He bet all
- 2:47his savings on the slowest horse in the
- 2:50race.
- 2:51A knight paid a priest to forgive a
- 2:53crime he hadn't yet committed.
- 2:56The paradox is literally in the very
- 2:58first sentence: "No intro, no greetings
- 3:00." And then, the tension immediately
- 3:02builds. Then there is a twist, where the
- 3:05real reason is revealed. And the final
- 3:07line brings the resolution. And they all
- 3:09have a signature technique.
- 3:1115 out of 20 videos transition to the
- 3:13explanation with the same two words,
- 3:15these ones.
- 3:16And they read this text at nearly four
- 3:18words per second. Meaning there are
- 3:20practically no pauses. The second layer
- 3:22is the visual itself. The scene's light
- 3:24falls onto the flat characters. They are
- 3:26tinted by the color of the place where
- 3:28they happen to be standing. The world
- 3:29casts real shadows, while the hero has a
- 3:31soft spot under their feet. Objects are
- 3:34outlined in black, like in a comic book
- 3:36. The colors are very saturated, with
- 3:39warm red-orange tones and cool, blue
- 3:41shadows. And most importantly, every
- 3:43scene has its own palette. Also, the
- 3:45camera barely cuts the frame. Thirteen
- 3:47out of twenty videos have only one cut,
- 3:49or none at all. The camera simply flies
- 3:51to the next moment itself. And the third
- 3:53layer, quite literally the most
- 3:55important. How did Claude break all this
- 3:57down? He didn't just eyeball it; he
- 3:59wrote himself a measuring tool. He
- 4:01turned every 2 seconds of video into a
- 4:03single image. 60 frames on one sheet.
- 4:06There are 577 such sheets.
- 4:08He calculated how much the image shifts
- 4:10from frame to frame. That is exactly how
- 4:12the camera itself moves. And he went
- 4:14ahead and marked every cut. He even
- 4:16measured the brightness, contrast, and
- 4:18saturation of every Short. According to
- 4:20these measurements, in 27%of the frames
- 4:22, the camera is literally standing in
- 4:24one place.
- 4:26Remember this number, because we'll come
- 4:28back to it.
- 4:29And here is the most important piece of
- 4:30advice: don't ask Claude what he thinks
- 4:32about a video. Just take it and make him
- 4:34measure it with some script. An opinion
- 4:37, for instance, cannot be repeated, but
- 4:38a number can. And that's it, our
- 4:40rulebook is ready. Now we make a similar
- 4:43Short, but with our own story. The same
- 4:45light, the same camera, the same cuts,
- 4:47and the same text formula repeated. Only
- 4:50the plot and the characters will be our
- 4:52own. My story is Clever Hans. Berlin,
- 4:561904. A horse that knows how to count.
- 4:58As long as someone who knows the answer
- 5:00is standing nearby. First, Claude
- 5:02checked the topic against all 147 shorts
- 5:04on the channel. They have a horse, but
- 5:06there are no overlaps in the other story
- 5:08. So, we kept the topic. Before writing
- 5:11the text, it verified 20 facts from the
- 5:13psychologist Pfungst's book, such as the
- 5:15date, the thirteen-member committee, and
- 5:17the fact that the horse was right 80%of
- 5:19the time without knowing the answer, and
- 5:2198%with it.
- 5:23Then it went ahead and wrote our script
- 5:24exactly according to the formula. A
- 5:26paradox in the first line. First comes
- 5:29the build-up, then the twist on those
- 5:31very two words, and finally the
- 5:33resolution. Next, Claude started
- 5:34building the video itself. And it builds
- 5:36it not from neural network video, but in
- 5:38a program that can generate 3D. This
- 5:40means frames assembled by code,
- 5:42literally right in our browser. And it
- 5:44can do exactly what we need. Flat
- 5:47characters in a real 3D world. And the
- 5:49Remotion program turns all of this into
- 5:51video, frame by frame. We couldn't
- 5:53download free 3D models, so Claude just
- 5:55drew everything with code. That is, the
- 5:57courtyard, houses, trees, and
- 5:59cobblestones.
- 6:00It made 10 faces for each of the eight
- 6:02characters, and eight separate ones for
- 6:04Hans the horse.
- 6:05From smiling and laughing to nervous and
- 6:08stunned. And the black outline, like in
- 6:10a comic book, it scripted itself; it
- 6:12finds the edges and traces them with a
- 6:13line. After half an hour, all 16 shots
- 6:15were working. Before rendering, Claude
- 6:18checked individual frames and found 17
- 6:20glitches where our characters went
- 6:22off-screen. A top hat covered a face,
- 6:24and the dust by the hooves looked like
- 6:25cotton wool. It took everything and
- 6:27fixed it. And this is our first version
- 6:29. I placed our frame next to the
- 6:31original. And, honestly, it looked very
- 6:33cheap. Then I told Claude directly. The
- 6:36original is bright, high-contrast, and
- 6:37looks expensive. But ours is dull.
- 6:39Analyze it frame by frame and don't rush
- 6:41. Use 10 agents if you have to, just
- 6:43make it as close as possible. And Claude
- 6:45agreed that ours was dull and cheap. Our
- 6:47shadows fall too deep into black. The
- 6:49camera barely moves and there is no
- 6:51lighting at all. And this is where the
- 6:52most interesting part begins. I asked
- 6:54for 10 agents, but Claude took six. An
- 6:57agent, it turns out, is a separate
- 6:58assistant that handles one specific task
- 7:00. Five of them each measure the
- 7:02difference in their area and redo their
- 7:04part themselves. That is, the camera,
- 7:06the world and lighting, materials,
- 7:08characters, and post-production. Each
- 7:10one puts the original frames next to
- 7:12ours and measures the difference. So it
- 7:14doesn't do it by eye, but literally with
- 7:16numbers. And the sixth one double-checks
- 7:19our own rulebook. But five agents in one
- 7:21project could break each other's work.
- 7:24Therefore, Claude gave each one their
- 7:25own copy of the project and their own
- 7:27files that no one else touches, and then
- 7:28assembled everything back together at
- 7:30the end.
- 7:31So, a tip for the future: when we have
- 7:33many agents, we shouldn't divide them by
- 7:35tasks, but by specific files.
- 7:37And here is what all our agents found.
- 7:40Our outline was too thick, about 11
- 7:42pixels, while the original is 2 to 6.
- 7:44Our image is also blurry. The original's
- 7:47sharpness is 676, while ours is 312.
- 7:51The camera was almost static in all 5
- 7:53cases, and the shadows were crushed into
- 7:55black. The original has almost none of
- 7:57that. And the agents also found errors
- 8:00in our own notes. Remember that 27%mark
- 8:02where the camera is positioned? That
- 8:05turned out to be untrue. It’s actually
- 8:07at 6 to 9%of the frames. It just crawls
- 8:10so slowly that the first tool literally
- 8:12didn't notice it. The build itself took
- 8:15Claude half an hour. The agent team's
- 8:17edits took a little over an hour. Here
- 8:19is what we ended up with. On the left is
- 8:21the 31 million original, and on the
- 8:22right is our video. And here is our full
- 8:25short.
- 8:39I’ve compiled everything from this video
- 8:41, every rule and every edit, into one
- 8:43skill for Claude. A skill is a
- 8:44ready-made instruction. You just need to
- 8:46put it into Claude, and it immediately
- 8:48knows the whole process. From breaking
- 8:49down clips to the finished video. It
- 8:51will be available for free on my
- 8:53Telegram channel. The link is in the
- 8:54description. You don't actually need to
- 8:56set up anything here. Just install it
- 8:58and make cartoons like this. And the
- 9:00main takeaway from this is: This process
- 9:02works not just with one style. Download
- 9:0420 clips, make Claude measure them, not
- 9:06just evaluate them. Put your frame next
- 9:09to the original. And this way, you can
- 9:11copy almost any style. Every week I
- 9:13break down formats like this on the
- 9:15channel. Subscribe so you don't miss out
- 9:17. And also don't forget to hit the like
- 9:18button. See you later.
About this transcript
This page contains the full transcript of Как Делать Вирусные Мультики-Шортсы через Claude (Полный Процесс) by ЗОЛОТЫЕ ЧЕКИ, generated from the public captions YouTube serves with the video. The transcript has 1,806 words across 272 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 choose a caption track available on the video.
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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.