Architectural knowledge and more specifically design deci- sions have become first-class entities to be captured routinely in a design process. In recent years, a number of approaches, tools, and formats have been proposed for handling the relevant design decisions for a system. However, the quality of the decisions captured is often low for many reasons. Part of the problem is that reflections intended to criticize, and thus improve, the decisions are seldom made, typically due to poor reflective practices in the architecture team. In this regard, some experiments have shown that poor reflections (e.g., by novice architects) produce low- quality decisions. In order to improve reflective practices and capture better design decisions, we propose an approach that integrates a design assistant for typical reflective tasks of the design process using generative AI techniques. Our assistant, called ArchMind, relies on two information sources: architectural knowledge about existing software patterns, and information about the system under design (e.g., context, requirements, past decisions). Furthermore, the assistant takes advantage of the capabilities of Large Language Models to progressively aid architects in the selection and analysis of alternative decisions, until capturing the chosen decisions in an Architecture Decision Record format. This work describes the assistant prototype and discusses initial results from a comparison with architects’ decisions from a classroom experience.