Break a formula into small steps and draw them as a graph. Compute forward to get the output, then walk backward: each node knows its local derivative, and multiplying it by the gradient arriving from above gives the gradient for its inputs. That is the chain rule, and it is all backpropagation is.