In the realm of architectural design, efficiently and effectively translating abstract design concepts into tangible 3D CAD models without extensive iteration remains a grand open challenge. At the core of this challenge lies the intricate task of capturing and manipulating the design reasoning process—a critical component yet to be fully addressed within existing digital design tools. We propose Compositional Neural Patterns (CNP) as a blueprint aimed at transcending these limitations by drawing upon the compositional nature inherent in architectural design—viewed through the lens of Category Theory, a mathematical tool grounded in compositionality. CNP is structured around three conceptual elements: design objects, design morphisms, and design compositions. Harnessing category theory’s power of abstraction, these elements enable the representation of design components, design logics and their interactions in a novel, computationally exploitable manner. This theoretical groundwork allows us to reimagine building blocks (design objects) and their interactions (design morphisms), laying out compositions that directly map thought-to-construction sequences otherwise mired in iterative cycles. While we are envisioning a deployment strategy that involves conjoining the theoretical foundation with deep learning, in this study we demonstrate CNP through a prototype plugin built in Grasshopper for Rhino3D. Specifically, we consider site, spatial organization grid, and massing as design objects, and demonstrate their representation, interactions (i.e. design morphisms), and design composition using Skidmore, Owings & Merrill’s (SOM) 1952 Lever House. We present our insights from this demonstration and reflect on future work.