Abstract: Examining large-scale, long-term application use is critical to understanding how an application meets the needs of its user community. However, there have been few published analyses of long-term use of desktop applications, and none that have examined applications that support creating and modifying content using direct manipulation. In this paper, we present an analysis of 2 years of usage data from an instrumented version of the GNU Image Manipulation Program, including data from over 200 users. In the course of our analysis, we show that previous findings concerning the sparseness of command use and idiosyncrasy of users' command vocabularies extend to a new domain and interaction style. These findings motivate continued research in adaptive and mixed-initiative interfaces. We also describe the novel application of a clustering technique to characterize a user community's higher-level tasks from low-level logging data.