Abstract: One or more computer processors derive a base corpus by capturing time series haptic data from an artisan performing an artisanal skill comprising a plurality of artisanal actions, wherein each artisanal action in the plurality of artisanal actions is associated with a time interval. The one or more computer processors train a multi-agent reinforcement learning (MARL) model utilizing the derived base corpus, wherein the MARL model outputs a vector representing one or more predicted artisanal actions required to complete the artisanal skill. The one or more computer processors generate a training avatar utilizing a trained generated adversarial network fed with one or more user parameters identified from a user requesting guidance for the artisanal skill and the output vector of the trained MARL model. The one or more computer processors guide the user with the generated training avatar in an execution of the artisanal skill.