This article presents an augmented reality-based instruction (ARBI) system for maintenance tasks. A traditional manual instruction method and a computer-assisted instruction method were compared. Three maintenance instruction methods, three task difficulty levels (low, medium, and high), and the user's gender (male and female) were specified as the independent variables in the experimental design. The dependent variables included task completion time and error rate as objective measures, and system usability scale (SUS) and NASA-task load index (NASA-TLX) scores as subjective measures. There were 30 participants (15 males and 15 females) in the experiment. The results indicated that the instruction method and task difficulty significantly affected the task completion time, error rate, SUS, and NASA-TLX.Among the instruction methods, the ARBI method exhibited the highest SUS score, lowest NASA-TLX score, shortest task completion time, and minimum error rate. In conclusion, the proposed ARBI method was beneficial for assisting iPhone maintenance tasks.