In the era of rapidly advancing smart manufacturing, product assembly is facing higher demands for flexibility and efficiency. The application of Augmented Reality (AR) technology in human-assisted assembly tasks has been increasingly prevalent, which can reduce cognitive load and enhance assembly efficiency. However, existing AR assisted assembly systems often neglect Expertise Reversal Effect, which refers to the phenomenon where different instructional strategies may have opposite effects on inexperienced learners and proficient learners during the learning process. In the case, skilled operators are provided with redundant information, resulting in decreased assembly efficiency. Based on this, this study proposes a proficiency-level grading model, considering user state to recommend guidance information in AR assembly tasks. Firstly, we divide the guidance information into different levels, facilitating the provision of adaptive information content. Secondly, we use HoloLens2 to gather user state across different skill levels and develop a proficiency-level grading model. Finally, leveraging the aforementioned research outcomes, an Augmented Reality assembly system is developed, enabling proficiency-aware differential guidance information delivery in the actual reducer assembly scenario. This study provides a solution to mitigate the negative impact of the Expertise Reversal Effect in Augmented Reality assisted assembly, aiming to improve the efficiency of the assembly process and enhance the user experience of the system. The proposed method offers new perspectives and approaches for the development and application of Augmented Reality Assistant System.