System management becomes increasingly complex and brings high costs, especially with the advent of Cloud Computing. In a Cloud, numerous platforms like Virtual Machines (VMs) and Middleware have to be managed to make the whole system work cost-effectively after an application is deployed. For controlling the management cost, in particular the manual management cost, many computer programs have been developed to take over manual management tasks or reduce their complexity and difficulty. These programs are usually hard-coded by languages like Java and C++, which bring enough capability and flexibility but also cause high programming effort and cost. This paper proposes an architecture based approach to developing the management programs in a simple but powerful enough manner. First of all, the manageability (such as APIs, configuration files and scripts) of a given platform is abstracted as a runtime model of the platform's software architecture, which can automatically and immediately propagate any observable runtime changes of the target platforms to the corresponding architecture models, and vice versa. Then the management programs will be developed using modeling languages, instead of those relatively low-level programming languages. Such architecture-level management programs bring many advantages related to the performance, interoperability, reusability and simplicity. The experiment on a real-world cloud and the comparison with the programming language approach demonstrate the feasibility, effectiveness and benefits of the new approach for management program development.