How to let AI agents sketch before painting?

Current image-generation agents are trapped in a loop of rewriting prompts to refine outputs, with no way to directly manipulate the result. GenClaw introduces a three-stage workflow: the agent reasons about the concept, then writes code (SVG, HTML, Three.js) to create a structural sketch, then uses image generation to add textures and realism. Code acts as a controllable bridge between language reasoning and pixels, giving the agent something closer to a real artist's toolkit—conceptualize, sketch, paint.