The paper examines the application of neural-network-based tools in the architectural design process, focusing on the Raven plugin, which generates and modifies Grasshopper definitions from natural-language prompts. The tool was tested through the conceptual design of a hotel within an artists’ residency in Gardinovci, developed for the Radical Innovation Hospitality Project competition. Results show that Raven performs more reliably when prompts describe parametric logic rather than the desired appearance or a reference image. As model complexity increases, the tool’s ability to maintain a coherent structure declines, requiring active designer involvement.