We describe a system that can learn new procedure models effectively from one demonstration by the user. Previous work to learn tasks through observing a demonstration (eg, Lent & Laird, 2001) has required observing many examples of the same task. One-shot learning of tasks presents a significant challenge because the observed sequence is inherently incomplete–the user only performs the steps required for the current situation. Furthermore, their decisionmaking processes, which reflect the control structures in the procedure, are not revealed. We will demonstrate a system called PLOW (Procedural Learning on the Web) that learns task knowledge through observation accompanied by a natural language “play-by-play”. Natural language (NL) alleviates many task learning problems by identifying (i) a useful level of abstraction of observed actions;(ii) parameter dependencies;(iii) hierarchical structure;(iv) semantic relationships between the task and the items involved in the actions; and (v) control constructs not otherwise observable. Various specialized reasoning modules in the system communicate and collaborate with each other to interpret the user’s intentions, build a task model based on the interpretation, and check consistency between the learned task and prior knowledge. The play-by-play approach in NL enables our task learning system to build a task with highlevel constructs that are not inferable from observed actions alone. In addition to the knowledge about task structure, NL also provides critical information to transform the observed actions into more robust and reliable executable forms. Our system learns how to find objects used in the task, unifying the linguistic information of the objects with the semantic representations of the user’s NL descriptions about them. The objects can then be reliably found in dynamic and complex environments. See Jung et al (2006) and Chambers et al (2006) for more details on the PLOW system.