Reinforcement learning algorithms powerfully explain how animals or artificial agents optimize their behavior based on environmental feedback. While neuronal correlates of such algorithms are observed in multiple subcortical regions, it remains unclear whether cortical computation can be guided by similar reinforcement feedbacks. Here, we show that cortical VIP-expressing inhibitory interneurons are recipients of a global and modality-independent reinforcement prediction signal conveyed by cholinergic inputs. This global reinforcement signal is then customized by local activity via a VIP and SST interneuron feedback loop, that is engaged upon learning. Using a computational model, we demonstrate that the recruitment of this VIP-SST interneuron motif by global cholinergic inputs enables reinforcement-dependent plasticity for learning of local features represented in cortex. Our results provide a circuit-based framework for adaptive reinforcement learning in the cortex, transforming global reinforcement signals to local teaching signals for improving task-relevant representations.