# Claude Code made a 3-minute animated music video entirely with code: 58 sub-agents, 38 hours. Process, numbers and what went wrong. — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/6139e56c7f6061f6

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-10-05 18:09 ingested 2026-10-06 02:37

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1wyfq5y/claude_code_made_a_3minute_animated_music_video/)

## retained text



```
Roughly 5 months of full time work for a professional animator/producer to produce something similar achieved in about a day and a half by Claude. I'm flabbergasted



\- Lyrics: written with Claude in chat. 

\- Music: generated with Suno from those lyrics.

\- Video: Claude Code (Opus 5.5) ran as an orchestrator and spawned sub-agents. Every frame is rendered from code with Remotion (React + TypeScript). There are no generated images or video in the picture.

\- Audio analysis, also run by Claude: Demucs for stems, beat\_this for the beat grid, wav2vec2 forced alignment for word timing.

\- Me: the brief, a review at the end of each phase, and notes.



**Process**



1. Audio analysis: stems, beat grid, and a frame number for every sung word.

2. Three look directions as style frames. I picked one.

3. The engine (camera, lettering, dance system) and one fully finished chorus as the reference scene.

4. The other 13 sections animated in parallel, one agent per scene folder.

5. Polish: an independent critic agent scored the cut, other agents worked through its list, then again.

6. 4K render, thumbnails, captions.



A dashboard (a Claude artifact) showed progress and collected my review notes at each gate.



**Numbers**



\- 38 hours wall clock from brief to 4K master, including the time it waited on my reviews

\- 58 sub-agents; about 54 hours of agent time, run in parallel

\- About 7,000 model responses and 10,000 tool calls

\- 5.6M output tokens; 2.6B tokens processed in total, nearly all of it cached context re-read on each step

\- About 45,000 lines of TypeScript; 11,478 frames at 4K 60 fps



**What went wrong**



\- Word timing. Forced alignment started 21 words late, mostly legato words where it locked onto the vowel. I heard one by ear; an audit against the vocal stem found the rest.

\- Camera. I asked for more camera movement and the first attempt was still timid (moves under 30 px). It only worked once agents were given numeric targets for travel, push and roll.

\- Self-grading. Agents checking their own scenes reported them clean. Separate checker agents, looking at the frame before and the frame of every sung word, found six real problems.

\- Lyrics appeared before they were sung, as outlines waiting to fill. Fixing it meant enforcing the rule in the shared lettering component first, then a pass over every scene.

\- A resumed sub-agent stalled silently. I noticed before the orchestrator did.

\- The first 4K render ran Chrome out of memory after 30 minutes and kept nothing. The second rendered in resumable parts.

\- Nothing in the pipeline watches the video play. Every automated check was on still frames and measurements; playback review was me.
```



document fetched 2026-10-06 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1wyfq5y/claude_code_made_a_3minute_animated_music_video
