Testing user interface and system-level functionality of a mobile app is crucial for ensuring its quality. However, it is becoming increasingly costly due to the complexity of modern applications and the diverse range of devices. Recent approaches have focused on exploring entire applications to test and detect defects in mobile apps. Additionally, they do not consider the ability to guide and restrict large language models (LLMs) based on user-defined rules. This paper introduces VisiDroid, an approach to generating scripts for mobile testing from task goals or natural language descriptions by leveraging the capabilities of LLMs. We evaluate the approach using an open-source dataset consisting of 131 tasks on 11 mobile apps. The results show that VisiDroid can accurately generate actions and achieves a task completion rate of 72.2%, outperforming the state-of-the-art approach. It also successfully generates valid test scripts with an 80.05% success rate overall.