Inside PDoomVideo: The Open-Source Music Video Rendered Frame-by-Frame in Code by Claude Opus 5.5
An architectural deep-dive into JohnHeibel/PDoomVideo, an open-source repository where Claude Opus 5.5 generated every frame and transition using p5.js, p5.brus
Rather than relying on diffusion-based black-box video synthesis models, a distinct engineering approach has captured attention across developer communities: instructing a large language model to write browser graphics code that paints every individual frame on an HTML canvas. The open-source project 'PDoomVideo' provides a concrete implementation of this paradigm, offering a complete, code-driven motion graphics pipeline.

Image source: JohnHeibel/PDoomVideo (GitHub)
Created by developer John Heibel and hosted at JohnHeibel/PDoomVideo, the repository serves as the complete source code for the animated music video set to the song "I'm Upping My P(doom)". As of November 13, 2026, the repository has gathered 2,922 stars and 281 forks on GitHub, demonstrating widespread interest in deterministic, programmatically rendered generative video.
Autonomous Generation Pipeline: From Minimal Directives to Codebase
What distinguishes PDoomVideo from conventional creative coding projects is that every visual scene, layout decision, and transition was generated directly by the model without manual storyboard drafting. The human author provided only two high-level constraints: utilize Anthropic's Clawd character design as the central protagonist, and generate visually compelling scene transitions synchronized to each lyric line.
The entire codebase emerged across two distinct generational passes using Claude Code:
- Generation 1 (archived under
legacy/): Built using Claude Opus 5.5 (Medium). - Generation 2 (current root codebase): Generated entirely by Claude Opus 5.5.
Before writing the low-level rendering scripts, the model orchestrated its own execution workflow. It first generated ANIMATION_GUIDE.md to serve as a design and formatting brief for the parallel subagents it spawned. It then formulated STORYBOARD.md, a comprehensive shot-by-shot visual sequence mapping out how each scene should evolve using p5.js brushstrokes and continuous geometric handoffs.
Architecture and the Headless Rendering Engine
Instead of producing lossy, hallucinated video clips where hands drift and background elements morph unpredictably, PDoomVideo operates through deterministic web canvas graphics. The visual foundation relies on two core JavaScript libraries: p5.js for canvas manipulation and p5.brush for textured, hand-painted digital strokes.
The repository's internal architecture is divided into targeted components:
studio.html: The runtime web page where every frame is actively painted onto the HTML5 canvas using p5.js and custom shader routines.src/ch/: Houses nine modular chapter files, each responsible for rendering the graphic logic of a specific song movement.src/Shared Module: Encapsulates reusable definitions, including the Clawd mascot, secondary characters, environmental props, synchronized lyrics, and the global audio timeline.render.mjs: A Node.js orchestration script that spins up headless Google Chrome instances to rasterize frames to PNGs, then executes ffmpeg to encode the frame buffer into a synchronized MP4 video.legacy/: Preserves the first-generation implementation for comparative analysis.
Because every visual element is defined through mathematical paths and programmatic state machines, the output maintains frame-to-frame coherence and razor-sharp line quality impossible to achieve through standard pixel-space diffusion generators.
Local Setup and Step-by-Step Rendering Workflow
Developers interested in inspecting or reproducing the render locally require Node.js, Google Chrome, and ffmpeg installed on their operating system. The original soundtrack is bundled directly at assets/pdoom.mp3.
Executing the render pipeline involves two deterministic CLI steps:
# 1. Install project dependencies
npm install
# 2. Paint frames into out/frames using 4 parallel workers (resumable)
node render.mjs --frames=0:156.6 --workers=4
# 3. Concatenate rendered frame buffers with audio into MP4 format
node render.mjs --encode --out=out/pdoom.mp4
For environments where Google Chrome is not located at the default system path on Windows, users can pass the --chrome=<path to chrome> flag to the runner. The --frames parameter enables precise segment rendering, allowing long runs to pause and resume without losing previously computed frames.
Ecosystem Context and Verified Scope
GitHub metadata confirms that JohnHeibel/PDoomVideo was originally created on September 22, 2026, with core pushes finalized on September 25, 2026. The project later went viral during community discussions in mid-October highlighting notable projects built during the rollout of Claude Opus 5.5, with curated threads such as Leonard Rodman's (@RodmanAi) listing it as a prime example of full-code generative video.
Following this experiment, the author also published JohnHeibel/ClaudeAnimationBase, a generalized and stabilized starter kit designed specifically for building multi-agent canvas animations with Claude Opus 5.5.
While PDoomVideo serves as a compelling proof-of-concept for code-based animation, its current codebase remains tailored to the specific timing and narrative arc of the "I'm Upping My P(doom)" track. It is not an arbitrary text-to-video desktop application, but rather an open architectural blueprint for developers looking to replace fragile diffusion pipelines with inspectable, version-controlled web graphics code.
Sources
- GitHub Repository: JohnHeibel/PDoomVideo
- Related Framework: JohnHeibel/ClaudeAnimationBase
- Leonard Rodman (@RodmanAi): 10 GitHub repos from Claude Opus 5.5 week