AI models are highly optimized for visual plausibility, but design intent lives in structure: constraints, usage context, and consistency across outputs. This becomes visible when the same prompt must produce variations that remain usable, not just visually correct. While building an image workflow system, this limitation kept surfacing — raising the question of whether intent can be modeled as something more than style, or if it requires an entirely different abstraction layer.