Big Idea

The argument in one line.

Claude Code makes distinctive animated videos only when you give it a numeric style skill and keep fixing that skill from your review notes, rather than rewriting prompts.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A faceless-channel creator who wants animated videos without animation software, stock footage or an image generator, and is tired of stick-figure AI slop.
A Claude Code user who can run commands in a folder and wants a repeatable pipeline from script to finished MP4 instead of one-off experiments.
A solo YouTuber who wants a consistent channel look that carries from video to video without redoing the prompt each time.

SKIP IF…

You want live-action editing or cutting existing footage; this workflow generates animation from code, not edits of recorded video.
You want a fully hands-off push-button video; the creator insists you stay the director and review every scene.

TL;DR

The full version, fast.

Videos built by Claude Code look like AI slop because nobody tells it how they should look. The fix is a workflow: let Claude set up Node, FFmpeg and Python, hand it a read-aloud script, generate a voiceover and run Whisper for word timing, then write a style skill that describes the look in numbers. Review a live browser preview instead of rendering, collect timestamped notes, and when notes repeat, update the skill rather than the video. Render in chunks optimized for your machine, then let Claude write titles and chapters. Start with 2D and a 20-second test.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:57

01 · Cold open: code-made video

Claims every frame was written by Claude Code and says AI slop comes from not telling it how.

00:58 – 01:44

02 · Setup and folders

Claude Code plus Node.js, FFmpeg and Python, installed by Claude itself, with one main folder structure.

01:45 – 02:41

03 · The script

Hand Claude the script, optionally with a camera line under each paragraph, and read it aloud first.

02:42 – 03:35

04 · Voice and Whisper timing

Text-to-speech options and Whisper word timing so animation moves on the words.

03:36 – 05:30

05 · The style skill

Why style matters, what a skill is, and describing the look in numbers.

05:31 – 06:17

06 · Sound and the 2D choice

Claude can make music and effects but over-loud; start with 2D to save tokens.

06:18 – 07:46

07 · Review and fix the skill

Live preview, timestamped notes, and updating the skill when notes repeat.

08:15 – 09:12

08 · Rendering and edit lists

Optimize and chunk the render, and reorder the timeline with an edit list.

09:12 – 10:17

09 · Upload and first steps

Titles, description, chapters and thumbnail, then the 20-second test challenge.

Atomic Insights

Lines worth screenshotting.

Claude has no style of its own; with no direction it falls back to the simplest thing it knows, which is why so many AI videos are stick figures.
A prompt says what to make, a skill says how everything you make should look, so a skill compounds across videos and a prompt is forgotten after one chat.
Describe a look in numbers, not adjectives: 'cinematic' means nothing, but a 24-degree lens and a slow push-in are rules Claude can follow.
Reviewing a live browser preview is instant because nothing is saved, while rendering a 13-minute film can take around 45 minutes of PC time.
When review notes repeat across scenes, fix the skill instead of the video; 152 notes on one film collapsed into a handful of recurring problems.
Claude cannot see its own visual bugs like trees spinning with the camera or people walking through furniture; only a human watching finds them.
Start with 2D, not 3D: 3D burns far more tokens, and moving cameras create clipping and continuity bugs.
Asking Claude to optimize the render for your computer cut multi-hour renders to roughly the length of the film.
Splitting a render into short chunks means a late fix only re-renders the chunks that changed.
An edit list of a few lines can reorder a timeline, such as moving the strongest moment to the start, without rebuilding any scene.
Claude tends to make sound effects too loud and samey, which is exactly how a video starts to feel like AI slop.
Start a new Claude session for every video to save tokens and let the skill carry what matters forward.
Whisper word timestamps let the animation move on the spoken words and let Claude retime scenes that are too fast or slow.

Takeaway

Make the style a skill, then keep improving it

WHAT TO LEARN

Distinctive AI video comes from a numeric style skill that you refine from review notes, not from clever one-off prompts.

Describe the look in measurable rules such as lens angle and camera motion, because adjectives like cinematic make Claude guess differently every scene.
Start from the script and a Whisper timing file so animation lands on the spoken words and can be retimed on request.
Review a live browser preview instead of rendering, so every note costs seconds, not a 45-minute render.
When the same note shows up across scenes, add a rule to the skill so every later video starts with that problem already solved.
Choose 2D over 3D at first, since moving 3D cameras cause clipping and continuity bugs and use far more tokens.
Ask Claude to optimize rendering for your machine and split it into chunks so a late fix only re-renders what changed.
Test with a 20-second scene three or four times and fix the skill each round to build a style that is yours.

Glossary

Terms worth knowing.

Skill: A folder with an instruction file that Claude reads before any matching job. It holds the rules for how everything you make should look and work.
Whisper: A free speech-to-text tool that records when each word is spoken, giving animations exact timing against the voiceover.
FFmpeg: A free command-line tool that stitches captured frames and audio into a finished video file.
Live preview: Playing the animation code in a browser in real time without saving frames, so you can review instantly instead of waiting for a render.
Edit list: A few lines of instructions that cut and reorder a timeline, letting you restructure a video without rebuilding its scenes.
Render: Drawing every frame of the animation and saving it into a video file, which is the slow step of the pipeline.

Resources

Things they pointed at.

00:58toolNode.js

00:58toolFFmpeg

02:45productElevenLabs

02:45productFish Audio

02:45productCartesia

03:09toolWhisper

06:26toolVideo Hub (creator's preview tool)

09:30productGoogle Flow

Quotables

Lines you could clip.

00:00

“Every frame you're about to see was written as code.”

strong cold-open claim, no setup needed→ IG reel cold open↗ Tweet quote

08:28

“A prompt is forgotten after one chat. A skill gets better with every video you make.”

tight contrast that lands alone→ newsletter pull-quote↗ Tweet quote

04:54

“Make it cinematic means nothing. Claude will guess, and every scene will guess differently.”

clear rule with a reason→ TikTok hook↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

Every frame you're about to see was written as code. A documentary about Hiroshima. A video about changing your body's settings.

No animation software. No stock footage. No image generator.

I didn't make any of these. Claude Code did. Meanwhile, YouTube gurus are selling the same dream.

Grab a stick figure style. Post every day. Make thousands a month.

So everyone did. And now YouTube is full of AI slop. That's not because Claude can't do better.

It's because nobody tells it how. So let's make one properly, from an empty folder to a finished video. We'll follow one 20 -second scene through every step, with the prompts I actually use, and bigger projects will show you where each step leads.

Let's start with the setup. Claude Code is Claude working inside a folder on your computer. It doesn't just chat, it writes files and runs them.

I use it with a paid Claude plan, and that's the only part of this setup that costs money. Everything else is free. Node .js to run the code, FFmpeg to turn frames into a video, and Python for the sound.

You don't have to install these by hand, by the way. Open Clawed Code in an empty folder and tell it, I want to make animated videos with code on this computer. Check what's missing and set it up.

It'll check what's missing and help you install it. You just approve the steps. One more setting.

Choose Opus 5 .5 as your model. It's the best at understanding context and visuals, and that matters a lot here. Then ask Claude to make one main folder for your videos, with subfolders inside.

One for finished videos, one for the engines it builds, one for your skills. It sounds boring, but after 10 videos, it's what keeps everything from falling apart. Next, the script.

every good youtube video starts with a good script if the script is bad no visual can save it but a good script usually turns into a good video so start with the words not the visuals the more your script describes what happens the better claude understands what to show and you have two options if you want the least work claude can make the whole video from one prompt or you go scene by scene and give claude your ideas for each one here's what i actually do i hand the script straight to claude and it plans the shots itself.

Opus 5 .5 is very good at this because it understands the context of every sentence. If you want more control, add one rule. Under every paragraph, one line describing what the camera sees.

So you get a sentence like, you're 15 and you work at the tram depot. And under it, the depot at dawn, trams in a row, a tag over you, age 15. Either way, read the script out loud.

cut anything that sounds like a textbook, and only then move on. Then, the voice. I use a text -to -speech voice, but if you can clone your own voice, it'll sound more original, and it's better for your channel's brand.

That part is up to you. Good paid options right now are Eleven Labs, Fish Audio, and Cartesia. For free, try Google's text -to -speech or Microsoft's Playground.

Either way, you end up with one audio file. Then ask Claude to run Whisper on it. Whisper is a free tool that writes down when every word is spoken.

Now, Claude knows exactly how long everything takes and the animation can move on the words. You can also ask Claude to add pauses wherever the film needs a moment to breathe. And if an animation ends up too fast or too slow for the voice, just tell Claude.

It'll retime it for you. Now, the style. This is the step that separates a good video from the slop.

Claude has no style of its own. give it no direction and it falls back to the simplest thing it knows that's where all the stick figures come from and the fix isn't a better prompt it's a skill and here's why style matters so much right now one channel typed something like make a 3d animation about something into claude uploaded the result and got a few million views within weeks hundreds of copycat channels appeared and almost all of them get nothing Style matters and branding matters.

Treat your channel like a business and a brand, not a lottery ticket. A skill is a folder with an instruction file in it. Claude reads it before any job that matches.

Your prompt says what to make. The skill says how everything you make should look and work. You write it once and every video, every scene, every new session follows it.

You don't have to write it from scrap. For reference, you can download videos or images from Pinterest and let Claude analyze them. Or, though I wouldn't recommend it, a part of a YouTube video you like.

The problem is, right now everyone is copying everyone, reverse engineering each other's videos. So originality is still the most important part. Why a skill at all?

Because every time I hand Claude a voiceover, I want the video to come out in the same style. That's really what you're building, a brand for your channel. A good skill describes the look in numbers, not adjectives.

Make it cinematic means nothing. Claude will guess, and every scene will guess differently. This skill, for example, asks for a high three -quarter camera with a narrow 24 -degree lens, still shots that slowly push in, and a handheld camera after a disaster.

Those are rules Claude can actually follow. It also says how the story is told on screen. This one says, show, don't tell.

no subtitles just little pixel name tags over people and objects and the characters themselves how they look and move claude works out from the context claude is really smart here if your test doesn't match your reference it can go through it frame by frame and find the differences but that burns a lot of tokens and we're trying to stay efficient that's also why i'd start with 2d not 3d 3d eats far more tokens and the jumping quality isn't worth it Claude can also make the music and sound effects encode, but be careful.

It tends to make the sound effects too loud, and they all sound alike, which is exactly how a video starts to feel like AI slop. You can always bring your own music and sounds instead. Next comes the fix.

This is where a video actually gets good, and it's the step almost everyone skips. Claude can make a whole video from start to finish, but someone still has to watch it, and the first thing Claude tries to do is render it. which can take a long time.

A 13 minute film takes my PC around three quarters of an hour. Do that 10 times, and that's seven and a half hours of just rendering. So I had Claude code me a program called Video Hub.

It has a live preview. Rendering means drawing every frame and saving it into a video file. The live preview just plays the code in the browser in real time.

Nothing gets saved, so it's instant. I watch the film scene by scene, and wherever something bothers me, I pause and leave a note. He's facing the wall, the tag covers his face.

Then all the notes go to Claude, it fixes those parts, and the preview updates. You don't need my tool for this, by the way. Pausing in a browser and writing down timestamps works too.

VideoHub just makes it faster. And if enough of you want it, I can put it on GitHub. And here's the rule that changed everything.

On my first full review of the Hiroshima film, I wrote 152 notes. Fixing them was easy, but many of them were the same problem in a different scene. So now, when notes repeat, I don't just fix the video.

I ask Claude to fix the skill. My skill keeps a list of those mistakes, and it now has 10 rounds of rules in it. Trees that spun with the camera, people walking through furniture, labels covering each other.

Every new film starts with all of that already solved. Claude can't see these by itself. They only show up when a person watches.

So expect them and put every one of them in the skill. And notice, these are all 3D problem. Once a camera moves through a 3D world, objects clip into each other and continuity mistakes pile up.

That's another reason to start with 2D. Fewer bugs, far fewer tokens. And honestly, Opus 5 .5 is excellent at motion graphics.

A prompt is forgotten after one chat. A skill gets better with every video you make. So, to save tokens, start a new session for every video.

The skill carries what matters from one video to the next. Then, the render. Only when the live version looks right do you render.

Claude opens the page in Chrome, captures it frame by frame, and FFmpeg turns the frames into a video with your voice and the music. This is where most of the bugs live. If you just say, render it, you might wait for hours.

Instead, ask Claude to optimize the render for your computer. The usual problems? The processor chokes, Chrome heats up and slows down, or a long render crashes near the end.

For me, renders that took several hours dropped to almost the length of the film itself. Splitting the render into short chunks helps too. If you find one more mistake later, you only re -render the chunks that changed.

And don't be afraid to fix the edit itself. My Hiroshima film first opened slowly, with the city waking up, and it didn't pull you in. So Claude wrote an edit list, a few lines that cut and reorder the timeline, and moved the most important moment to the very start.

A new structure. without rebuilding a single scene. And finally, the upload.

Claude already knows your script, so let it write the title options, the description, and the chapters. Thumbnails are a whole topic of their own, and honestly, this part is more about you than about Claude Code. If the title and thumbnail aren't good, it's very hard for any video to get views.

If you have something like a Google Flow subscription, Claude can write the prompt and pick the reference images, and you generate the thumbnail there. the film itself is all code the thumbnail is the one place an image model helps so that's the whole thing it's automation but not the push button kind you stay the director claude does the work so here's your first step open claude code describe a look you love and ask it to turn that into a skill give it a short script and ask for a 20 second test not a full film watch it write down what's wrong and fix the skill not just the video Do that three or four times and you'll have a style that's yours and nobody else's stick figures.

And if you want to see how the next one gets made, subscribe.

The Hook

The bait, then the rug-pull.

Every frame you see was written as code, no animation software, no stock footage, no image generator. The creator argues the slop flooding YouTube is not a Claude limit, it is a missing instruction.

Frameworks