While Large Language Models (LLMs) have significantly advanced design ideation and problem-solving, a fundamental tension persists between the discrete nature of current LLM interactions and the iterative, non-linear essence of design thinking. This “human-LLM misalignment” often forces designers to choose between accepting opaque “black-box” outputs or restarting the generation process entirely, thereby hindering the evolution of design intent and diminishing critical reflection. Based on a formative study with eight designers, we identify three core challenges in achieving reasoning-level alignment and propose a “Curated Reasoning” interaction approach. To instantiate this approach, we developed DesignerlyLoop, a prototype that introduces a nested two-layer diagram structure. This system allows users to externalize their evolving design intent while simultaneously inspecting, reorganizing, and selectively regenerating the underlying LLM reasoning chains. A within-subject user study with 20 designers demonstrated that curated reasoning significantly enhances design intent formulation, and output quality. By shifting the user’s role from a passive recipient of generated content to an active curator of AI reasoning, DesignerlyLoop fosters a more reflective and iterative human-AI collaborative process. Our contributions include: identifying key alignment challenges in creative design; the implementation of a dual-layer structure for reasoning curation; and empirical evidence validating how explicit curated reasoning supports more effective human-LLM alignment in creative tasks.