We revisit the standard hybrid CP/SAT approach for solving disjunctive scheduling problems. Previous methods entail the creation of redundant clauses when lazily generating atoms standing for bounds modifications. We first describe an alternative method for handling lazily generated atoms without computational overhead. Next, we propose a novel conflict analysis scheme tailored for disjunctive scheduling. Our experiments on well known Job Shop Scheduling instances show compelling evidence of the efficiency of the learning mechanism that we propose. In particular this approach is very efficient for proving unfeasibility.