How to Make Viral Animated Shorts Using Claude (Full Workflow) — Transcript
Full transcript
- 0:00Look at these two videos. This one has
- 0:0229 million views and it's not a one-off.
- 0:05That channel is brand new and its last
- 0:0780 shorts got over 200 million views.
- 0:10The other one, that's mine. And no, it
- 0:12didn't come out of an AI video
- 0:14generator. Claude built it in 3D with
- 0:17code every single frame. In this video,
- 0:19I'll show you exactly how I made this
- 0:21animated short, how I trained Claude on
- 0:23this style frame by frame. What went
- 0:26wrong and how I fixed it, and how you
- 0:28can get it, too. You're going to get all
- 0:30of that for completely free. All right,
- 0:33let's get into it. If you've never seen
- 0:35this type of short, it's this. The
- 0:37characters are 2D, but the world around
- 0:39them is 3D. So, how do you train Claude
- 0:41to make this? First, you need
- 0:43references. A lot of them. I went to the
- 0:46channel and had Claude download the last
- 0:4720 shorts in this style, the ones that
- 0:49look this good. Now, here's the thing.
- 0:51If you just tell Claude to analyze a
- 0:53video, it won't look at every frame. It
- 0:55grabs a few here and there, but a normal
- 0:57video runs at 30 frames per second. And
- 0:59you've probably seen this style all over
- 1:01your feed. Most of those look trash. The
- 1:04reason is the tiny stuff. Micro
- 1:06animations, a little camera move, a
- 1:08little shading, a little light. That's
- 1:10what makes it feel real. And you only
- 1:12catch that at 30 frames per second.
- 1:14Every single frame, every camera move,
- 1:16every little change in the light. So
- 1:18that's the main thing I told Claude.
- 1:19Watch all of them at 30 frames per
- 1:21second. and check everything. The
- 1:23lighting, the contrast, the shadows, the
- 1:26movement, the effects, the cuts,
- 1:28everything, 20 shorts, about 18 minutes
- 1:30of video, over 32,000 frames, and it
- 1:33came back with a rule book. So, what's
- 1:35the actual workflow behind these shorts?
- 1:38It comes down to three layers. How the
- 1:40script is engineered, how the visuals
- 1:42are constructed, and how Claude measured
- 1:44all of it. First, the script. To decode
- 1:46it, Claude needed the audio. And the
- 1:48first time YouTube handed it the German
- 1:50dub classic. So it pulled the original
- 1:53English track and transcribed all 20
- 1:55itself. And every single one opens like
- 1:58this. They buried a man alive with a
- 2:00bell tied to his finger. He bet his life
- 2:02savings on the slowest horse in the
- 2:04race. A knight paid a priest to forgive
- 2:06a crime he hadn't committed yet. There's
- 2:08a button in the middle of town and
- 2:10nobody's allowed to press it.
- 2:11>> A paradox in the very first sentence. No
- 2:14intro, no setup. Then it escalates. Then
- 2:17comes the pivot where the real reason
- 2:19comes out. Then the payoff. One last
- 2:21line and the twist lands. And there's a
- 2:23signature move. 14 of the 20 flip into
- 2:26the explanation with the exact same two
- 2:28words. You see? You see? You see? You
- 2:31see? You see?
- 2:32>> And the delivery is relentless. Almost
- 2:34four words a second with basically zero
- 2:36dead air. Second, the visuals. The
- 2:39characters are flat, two-dimensional,
- 2:40but they're lit like they physically
- 2:42exist in that world. They catch the key
- 2:44light and they cast real shadows. The
- 2:46world is real 3D geometry traced with
- 2:49black ink outlines like a comic book.
- 2:51The colors are heavily saturated, warm
- 2:53highlights, cool blue shadows, and the
- 2:55lighting follows the story. Warm and
- 2:57bright while it sets things up, then a
- 2:59deep night blue right as the explanation
- 3:01begins. And the camera almost never
- 3:03cuts. 12 of the 20 have one hard cut or
- 3:06none at all. Instead, the camera
- 3:08physically travels into the next moment.
- 3:10Third, and this is the real secret, how
- 3:12Claude analyzed all of it. It didn't
- 3:14just eyeball it. It wrote its own
- 3:16measuring tool. Every 2 seconds of video
- 3:18became one contact sheet. 60 frames on a
- 3:21single image, 549 of them. Then it
- 3:24tracked the camera with phase
- 3:25correlation, measuring how far the image
- 3:27shifts frame by frame. It flagged every
- 3:30hard cut. And it measured the
- 3:31brightness, the contrast, and the
- 3:33saturation of every short. That's how it
- 3:35found that in about 80% of the frames,
- 3:37the camera is almost perfectly still.
- 3:39Just tiny drifts. So, here's the tip.
- 3:42Don't ask Claude what it thinks of a
- 3:43video. Make it measure. You can't
- 3:45recreate an opinion. You can recreate a
- 3:48number. So now I had the whole
- 3:49blueprint. Now the story. We made the
- 3:52same type of short but with our own
- 3:54story. Same lighting, same camera, same
- 3:57cuts, same script formula, just our own
- 3:59characters. My first topic was why time
- 4:02seems to slow down when you're in
- 4:03danger. But while Claude studied the 20,
- 4:06it noticed one of them already covered
- 4:07the exact same experiment with 5.7
- 4:10million views. So, we threw it out and
- 4:13picked a new story, one the channel had
- 4:15never told. 1831, 74 soldiers march
- 4:18across an iron bridge, all in step, and
- 4:20one end of the bridge drops into the
- 4:22river. Before writing a single word,
- 4:24Claude checked every fact, the date, the
- 4:27number of soldiers, even how deep the
- 4:29river was, 2 feet. And that turned into
- 4:31the best joke in the whole short.
- 4:33Luckily, the river was only 2 ft deep.
- 4:36Unluckily, the river was only 2 ft deep.
- 4:39Then it wrote the script with the exact
- 4:41formula from the 20. A paradox in the
- 4:44very first line. 74 soldiers knock
- 4:46[music] down an iron bridge just by
- 4:48walking across it. It's then it builds
- 4:50up. Then you see then the twist. After
- 4:53that, Claude started building. At first,
- 4:56it used Blender. Same as my very first
- 4:58test. Before any of this, I asked Claude
- 5:01for a quick 5-second test just to see
- 5:03the look. Claude designed its own
- 5:04character, a little courier. It grabbed
- 5:07free 3D buildings, made flat cartoon
- 5:09clouds, and matched the camera timing
- 5:11from the channel's frames. There was
- 5:13just one problem. On the third camera
- 5:15push, the camera flew right past the
- 5:17character. So, for the bridge, Claude
- 5:19started in Blender again. It tested the
- 5:21ink lines there and drew the whole cast
- 5:23in code. Soldiers, an officer, people
- 5:25from the town, a kid, and a parent. Then
- 5:28I had a hunch. Maybe these shorts are
- 5:30made with 3JS. That's 3D built with code
- 5:33right in the browser. Claude checked the
- 5:35frames. Flat characters standing in a
- 5:37real 3D world. 3JS can do exactly that.
- 5:40So, it moved everything over and
- 5:42Remotion turns it into a video frame by
- 5:45frame. Then, it built the world. 37 free
- 5:483D models. Trees, rocks, flowers, market
- 5:51stalls, a cart, even a windmill. And
- 5:54every texture, the grass, the
- 5:56cobblestones, the roofs, the water was
- 5:58painted by code. The characters got 14
- 6:00different faces. laughing, whistling,
- 6:03nervous, shocked. Then the bridge. Two
- 6:06towers, iron chains, and a deck 5 m
- 6:08wide. 74 soldiers march across it, four
- 6:11side by side. Every step makes it
- 6:13bounce, and the bounce keeps growing
- 6:15until it snaps. For the ink lines,
- 6:17Claude wrote its own effect. It finds
- 6:19the edges in the 3D world and draws a
- 6:21black line on them. About an hour after
- 6:23the switch, all 15 shots were working.
- 6:262,25 frames. But before rendering any
- 6:29video, Claude checked still frames. Good
- 6:31thing it did. In three shots, the camera
- 6:34was stuck inside a house. The trees were
- 6:36teal and a chain was right in front of
- 6:38everyone's faces. While Claude was
- 6:40fixing that, I put one of our frames
- 6:41next to the original, and it looked
- 6:43cheap. Here's what I told Claude. The
- 6:46original is colorful, high contrast. It
- 6:48looks premium. Ours looks dull. So, go
- 6:51analyze it frame by frame. Take your
- 6:53time. Use 10 agents if you have to. Just
- 6:56make it as close as you possibly can.
- 6:58And Claude agreed. Ours looked cheap.
- 7:00Then it broke down exactly why. The
- 7:03original has black ink lines on the
- 7:05whole 3D world. Ours had almost none.
- 7:07Their colors are bold with long blue
- 7:10shadows and deep darks. Ours were pastel
- 7:12and flat. Their wood and stone look hand
- 7:15painted. You can see the grain and the
- 7:17dark gaps between the planks. Ours were
- 7:19plain. Their backgrounds are blurry so
- 7:21your eye goes straight to the character.
- 7:23Ours were mostly sharp. And the
- 7:24characters, theirs have egg-shaped heads
- 7:26and thin little limbs. Ours had perfect
- 7:29circle heads and chunky bodies. They
- 7:31looked like clip art pasted on top. And
- 7:33this is where it gets interesting. I
- 7:35asked for 10 agents. Claude used 12. And
- 7:38it didn't just throw them at the
- 7:40problem. It built a team. Seven agents
- 7:42to study the gap. One for color, one for
- 7:44the ink lines, one for light and
- 7:46shadows, one for textures, one for the
- 7:49characters, one for depth and framing,
- 7:51and one for motion. The camera and the
- 7:53characters. Each one put the original
- 7:55frames next to ours and measured the
- 7:57difference. No guessing, real numbers.
- 7:59An eighth agent worked on the sound.
- 8:01Then four more agents to rebuild
- 8:03everything. One for the camera, one for
- 8:05the world, one for the characters, and
- 8:07one for the final look. But here's the
- 8:09problem. Four agents working on the same
- 8:12project can break each other's work. So
- 8:14Claude gave each one its own copy of the
- 8:16project, and each agent owned its own
- 8:19files. Nobody else could touch them.
- 8:21When they were done, Claude merged
- 8:22everything back file by file. So, here's
- 8:25the tip. When you use a lot of agents,
- 8:27don't just split the jobs, split the
- 8:29files. Then, I gave Claude 10 hours. 10
- 8:32hours to learn this style and make the
- 8:34finished video. Here's what the seven
- 8:36agents found. First, the ink lines. The
- 8:39original has black lines on almost three
- 8:41out of four edges. On ours, the agent
- 8:44found zero. The ink effect was there,
- 8:46but the blur effect ran in the same step
- 8:48and wiped it out. And the blur itself,
- 8:50it was so weak you couldn't see it
- 8:52anywhere. Next, the camera. Ours snapped
- 8:54from one spot to the next in three
- 8:56frames. The original glides over about
- 8:5923 frames. Slow in, slow out. Then the
- 9:02color. The original has no pure black
- 9:04anywhere. Ours went almost black. And at
- 9:07sunset, their shadows are purple. Ours
- 9:09were brown. And the characters, their
- 9:11bodies are 3 and 1/2 heads tall. Ours
- 9:13were under three. And our arms and legs
- 9:16were five times too thick. But the
- 9:17agents also found something I didn't
- 9:19expect. Some of our own notes were
- 9:21wrong. Remember the egg-shaped heads?
- 9:23The agent measured 38 of them. Their
- 9:25heads are actually rounder than ours.
- 9:27Ours were the egg-shaped ones. And
- 9:29remember the measuring tool from the
- 9:30start? It said the camera is almost
- 9:32still 80% of the time. It's not. The
- 9:35camera is moving in 65% of the frames,
- 9:38slowly pushing in. The tool just
- 9:40couldn't see moves that slow. And this
- 9:42is exactly the part other tutorials
- 9:44skip. I found only two other tutorials
- 9:46on this style. One sells it inside an
- 9:48app, $67 a month or $97 for everything.
- 9:53The other one gives you a few prompts
- 9:55and it says, "Your first results might
- 9:57not look as clean." So, here's how you
- 9:59can get mine. Everything from this
- 10:00video, every rule and every fix I save
- 10:03as one skill for Claude and I'm giving
- 10:05it away fully free on my GitHub. You get
- 10:08the whole repository with 5,000
- 10:11reference images. The link is in the
- 10:12description. So, you don't have to do
- 10:14anything. Just install it and start
- 10:16making videos. Everything is already set
- 10:18up, so the output should be as polished
- 10:20as it gets. And I'm going to keep
- 10:22uploading new things there. I'll keep
- 10:24updating it and making it better. And
- 10:26you get all of that for free. And here's
- 10:28one more thing. I'm making a video on it
- 10:30tomorrow. This realistic 3D video, as
- 10:33you can see, I made it. That's my name
- 10:34on it. Here's something wild. Pull every
- 10:37blood vessel out of your body and lay
- 10:39them in one straight line. Most people
- 10:41guess it would reach across a football
- 10:43field. Not even close. Your body hides
- 10:46billions of microscopic tubes branching
- 10:48through every muscle, every organ, even
- 10:50your [music] bones. Lay them all out and
- 10:52the line would leave your city, cross
- 10:54entire countries, stretch over oceans,
- 10:57and keep on going. [music] Scientists
- 10:58estimate about a 100,000 km. That's
- 11:01enough to wrap around the whole Earth
- 11:03roughly 2 and 1/2 times. Just so you
- 11:05know, I've never seen anything like this
- 11:08made with AI. Other people have made
- 11:09this style with Claude, but not like
- 11:11this. I looked and I couldn't find a
- 11:13single tutorial that gets this close. If
- 11:16this showed up on your feed, you'd never
- 11:18guess AI touched it. In that video, I'll
- 11:20show you exactly how I made it. How to
- 11:22animate a realistic face with an
- 11:24open-source AI model, how to rig the
- 11:26character for free on your own computer,
- 11:28how to make the world look real, how to
- 11:30cut the render from 7 and 1/2 hours to 1
- 11:33hour, and how to render it smaller then
- 11:35upscale it so it renders as fast as it
- 11:37gets. You can find the skill and
- 11:39everything in the description. The link
- 11:41to the GitHub repository is right there,
- 11:43completely free. That's all for today.
- 11:45Make sure to subscribe and like. And
- 11:47with this type of workflow, you can
- 11:49pretty much clone any style. Bye.
About this transcript
This page contains the full transcript of How to Make Viral Animated Shorts Using Claude (Full Workflow) by Mr omi, generated from the public captions YouTube serves with the video. The transcript has 2,284 words across 323 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.