Different training conditions can have a profound influence on human performance, learning, and retention. Inspired by recent neuroimaging and modeling studies on the geometry of task representations, we explored how the sequencing of stimulus versus task information (context), and the use of identical versus distinct response labels across tasks, could influence task learning and accuracy. We hypothesized that presenting stimuli before contextual cues and employing distinct response labels for each task would foster more high-dimensional task representations, thought to improve retention of task knowledge. We evaluated multi-task learning in four groups (Context-first with/without Response labeling, and Stimulus-first with/without Response labeling) where participants first learned through trial-and-error to categorize stimuli differently depending on three different contexts. During a test phase without feedback, participants continuously switched between the three tasks. Accuracy significantly decreased from training to test, notably influenced by Response labeling, but not Presentation order. Task-specific response labels helped retain task information better, particularly in the stimulus-first group. We conclude that employing response labels and presenting stimuli first has the potential to improve task representations and multitasking accuracy. In a subsequent study, we explored whether learned task information could also be repurposed more efficiently to entirely new tasks, depending on initial Response labeling. Indeed, subjects that learned better with response labeling also performed better in a new, value-based decision-making task. Together, our findings highlight how different training interventions can influence the adaptability and transferability of task knowledge in a multi-task environment.