# I copied Disney's 1930 animation pipeline into a Claude Code workspace, and it makes my videos — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/7ff2c70479bd634e

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-09-30 00:50 ingested 2026-10-01 02:44

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1wtra8s/i_copied_disneys_1930_animation_pipeline_into_a/)

## retained text



```
Every screen in this video is a real file from the workspace that produced it: its brief, script, transcript, storyboard, scene code, stills and render. Here's how the workspace is structured, since that's the part I think carries over to anything. 

Claude Code can rip through it on sonnet and still make great video one at a time or throw on ultra code with Opus 5.5 and make 50 in a couple hours. 

All of this is based on my folder structure methodology called ICM. [https://github.com/RinDig/icm-architect](https://github.com/RinDig/icm-architect)

**The layout**

    animations/
      CLAUDE.md          <- the map: what lives where, where to go for each job
      CONTEXT.md         <- the whole pipeline in one table
      anim.py            <- every command (speak, transcribe, cues, stills, render)
      stages/
        01_voice/CONTEXT.md
        02_transcript/CONTEXT.md
        03_cues/CONTEXT.md
        04_scene/CONTEXT.md
        05_render/CONTEXT.md
      videos/<slug>/     <- one folder per video, 01_ to 05_ are its stage outputs

**Each stage has a small contract.** Here's the real one for the storyboard stage, trimmed:

    # 03_cues: the storyboard
    **Reads:** 02_transcript/transcript.md, 01_voice/brief.md, _shared/kit.md
    **Don't read:** words.json (the command does)
    1. Write 03_cues/cues.md, one cue per visual change
    2. python anim.py cues <slug> matches every phrase to the word timings
    **Output:** cues.md (source), cues.js + cues.json (generated, never edit)
    **Person checks:** reads the storyboard: every beat worth seeing, the story in the right order.

Reads, output and person checks. That's the whole pattern. Claude only loads the files a stage needs, and it doesn't move to the next stage until I've read what the last one made.

**The pipeline**

1. **Voice:** a script written the way I talk, voiced with my ElevenLabs clone (the video says so up front)
2. **Words:** Whisper gives every word a start and end time and flags anywhere the audio and script disagree
3. **Cues:** a plain text storyboard where each beat hangs off a few spoken words, so the timing comes from the voice
4. **Scene:** Claude writes the animation as JS from a kit we've built up (characters, rooms, props, real-screen pieces). Everything on screen is a function of time, with no timers. It shoots a contact sheet of every beat and reads it to fix overlaps.
5. **Render:** headless Chrome steps through each frame, then ffmpeg mixes the voice, music and effects

**What I learned**

* The contract's "Person checks" line matters more than anything else in it. A person reading each stage's file before the next one runs is what stops one bad take from turning into a bad video.
* Put the storyboard in a text file. Fixing a beat there takes seconds; fixing it in a rendered scene takes an hour mistakes in the script compound expodentially if you are automating without being able to see anything. 
* CLAUDE.md stays short. It routes Claude to the right folder, and the detail lives in each stage's CONTEXT.md, so each stage only loads the context it needs.
* Nothing here is Claude-specific. Any AI that reads files could walk the same folders.

The structure is borrowed from how Disney worked around 1930: voices recorded first, dialogue broken down frame by frame on an exposure sheet, rough animation screened in the "sweatbox" before anything got inked. The video covers that too.

Full walkthrough (7 min): [https://www.youtube.com/watch?v=EhWlGingCl0](https://www.youtube.com/watch?v=EhWlGingCl0)

Happy to go deeper on the contracts, the kit, or how the render stays frame-accurate.
```



document fetched 2026-10-01 00:44 · canonical https://reddit.com/r/ClaudeCode/comments/1wtra8s/i_copied_disneys_1930_animation_pipeline_into_a
