In this paper, we address flexible assembly plans generation to accommodate human task variability. Unlike existing approaches, the proposed approach is based on a free-style human-robot interaction (HRI) that does not impose any task order on the participants and furthermore can accommodate their errors and help them to correct the errors without stopping the whole assembly process. Our approach is implemented in a real robotic architecture that combines sensory-motor modules with Hierarchical Task Networks (HTN) to endow the cobot with the necessary adaptability to correspond to human actions dynamically. We show experimentally with 56 participants on a simulated industrial assembly task that the cobot increases the task performance (reduction in the number of errors and gestures) without increasing the participants' cognitive load.