AutoShader

— — —

Challenges

Loading…

Current Frames

No current frames yet — run an experiment to see results here.

Best Frames

No frames yet — run an experiment to see results here.

Best score: 0/100

Rubric

Experiment History

Time	Score	Best?	Comment	Failure Patterns

Current Shader (current/shader.glsl)

Loading retrospective…

Twitter / X Post Draft

Built an autonomous shader research loop: an AI agent that teaches itself GLSL by iterating through 18 progressive challenges, from basic 2D coordinates to full raymarched scenes.

How it works:
- Agent writes a shader
- Renderer produces frames (moderngl/EGL, no GPU needed)
- A vision model scores the output against a rubric
- Keep if improved, revert if not
- Repeat — skill doc evolves with each cycle

Results from the first run (Claude Opus as both agent + evaluator):
- 16/18 challenges completed in ~4.5 hours
- 172 total experiments
- Went from nothing to soft shadows, ambient occlusion, domain-warped SDFs, and full animated scenes
- Got stuck at 80/100 on the capstone challenge — the evaluator anchors to its own previous best and refuses to score higher

Interesting findings:
- Compositional coherence matters more than technique count. A simpler scene scored 90 while a more complex one plateaued at 80
- Domain warp amplitude of 0.17 is a magic number — higher causes raymarching artifacts, lower looks boring
- The biggest blocker wasn't the shader writing — it was the evaluator creating a local optimum trap by comparing side-by-side with the best

Next: swapping in a different eval model to see if it breaks the ceiling.

All open source — agent loop, renderer, evaluator, web dashboard, everything.

Auto-updated every 5 experiments