The Watson system automatically retrieves useful information in the context of a document-manipulation task. Previously-reported laboratory studies show the basic Watson system can consistently retrieve useful documents. However, a number of deployments in real-world settings suggest improvements in retrieval performance can be made. We describe a novel set of techniques for automatically transforming representations of a user’s context for the purpose of just-in-time information retrieval. The techniques are grounded in the experience gained during pilot deployments of the system. The methods described allow the system to identify salient properties of text the user is manipulating. This and additional representations of the user’s context are then used to guide a set of reusable query formation strategies, allowing the system to deliberately retrieve information of known relevance in the context of specific classes of tasks.