Back to browse
GitHub Repository

A DSL for generating 2D animated movies entirely from code. Procedural characters, custom poses, body angle perspective, walk cycles, camera systems — no external assets needed. Text in, video out. Built as a vibe coding experiment: every line written by Claude Opus 4.6, zero human code.

2 starsRust

A 2D animation DSL – every line written by Claude, zero human code

by siddhartha98·Feb 22, 2026·1 point·0 comments

AI Analysis

●●SolidWizardryBold BetZero to One

AI-written animation DSL with procedural bezier curves and spring physics, but output is still cartoon-level.

Strengths
  • Claude built 3,300-line procedural character renderer with zero human touches — genuine constraint satisfaction.
  • Iterated from rigid SVGs to procedural with body angle perspective and physics — real creative technical progression.
  • Constrains to agent-only code maintains experimental integrity — anthropologically interesting project.
Weaknesses
  • Output quality is limited to procedural cartoon faces — no path to photorealism or professional animation.
  • Narrow use case: existing animation tools (Blender, After Effects) already handle this with human artistry.
Target Audience

AI researchers, animation enthusiasts, developers exploring AI agent capabilities

Similar To

Blender · After Effects · Spine (skeletal animation)

Post Description

I wanted to see how far vibe coding could actually go. Not autocomplete, not copilot — I gave Claude Opus 4.6 (running in OpenCode, an agentic CLI) full ownership of building a domain-specific language for generating 2D animated movies from code. The constraint: I describe what I want, the agent builds it. I critique the output, the agent iterates. No human touches the code at any point. Not the parser, not the renderer, not the 3,300-line procedural character drawing module. Not even the demo scripts or SVG backgrounds. It went through three major rewrites: - V1: Rigid SVG characters sliding around. Puppet show. - V2: Bone-based rig system. Still puppet-like. - V3: Fully procedural — characters drawn frame-by-frame from bezier curves with body angle perspective, staggered joint timing, and spring physics on hair/clothing. The most interesting parts of the process: 1. The agent kept over-engineering infrastructure while the actual visual output stayed at stick-figure quality. It took direct, blunt feedback ("this looks like a puppet show, you keep building the same thing") to break the pattern. 2. Poses were originally 9 hardcoded words in Rust. After I pointed out this was a limitation, the agent added a full custom pose system to the DSL — 23 controllable joint/expression parameters defined in the script file. 3. The agent added overlap detection after characters kept walking through each other. It now refuses to render if two characters would intersect at any point. The output still looks like what it is — procedurally generated cartoon figures drawn with math. But 7,665 lines of Rust, a PEG grammar, procedural character rendering, camera systems, scene transitions, and a working demo — all from an AI agent that never saw a reference implementation. Repo policy: only agent-generated code accepted. Bug reports welcome, but all implementation must come from agents. Tech: Rust, tiny-skia for 2D rendering, pest for parsing, FFmpeg for video encoding.

Similar Projects

AI/ML●●Solid

Agent skill for Excalidraw diagrams with animation and image

Auto-animated Excalidraw diagrams from prompts beats manual canvas setup.

Ship ItSolve My Problem
muthuishere
102mo ago