vibes - spec canvas

Thesis: you can’t prompt taste, but you can curate it. Rather than asking users to articulate what they want in words, Spec Canvas lets them point at something tasteful and remix it – with AI generating the interface itself, not just the output.

The working prototype is a live remix workbench built around a WebGL shader (a “Nova Halo” component, originally extracted from Framer AI’s halo-glow component). Two Claude agents negotiate behind the scenes: Alice analyzes the shader’s documentation and extracts a taste-to-math mapping - translating 12 shader uniforms into perceptual categories like “Edge Turbulence” and “Color Warmth.” Bob then generates a complete UI control schema via structured tool use, deciding what controls to expose, how to group them, and what to call them. The controls you see on screen weren’t designed by us. Claude decided what knobs a non-technical person would need to explore this artifact.

The interface shatters conventions in three ways. First, the UI is generative: different artifacts would produce entirely different control panels. The sliders, color pickers, and groupings are Claude’s output, not a static form. Second, direct manipulation and natural language coexist; drag a slider at 60fps for precise control, or type “make it feel like deep space, cold and lonely” and watch Claude interpret that into coordinated parameter shifts across hue, palette, and noise. Third, the A2A (agent to agent) trace panel makes the AI collaboration visible: expand it to see Alice’s extraction reasoning and Bob’s parameterization decisions, with syntax-highlighted JSON and timing badges.