The trade-off you want to make is: (1) how hard is it to solve the master problem (and hence, how much time does it take), and (2) how many Benders iterations does it take to find a feasible/optimal solution (depends on what you are looking for). This you can best determine through some benchmarking.
As a rule-of-thumb, you want to add a relaxation of the subproblem to the master problem, to guide the master problem in the right direction. The trick here is to keep the master problem light enough that it can be solved efficiently.
A standard example is the capacitated facility location problem where the master problem determines which facilities to open while minimizing opening costs, and the subproblem solves an assignment problem to assign customers to opened facilities while minimizing assignment costs. Since the master problem attempts to minimize the total cost to open facilities, in the first iteration, in the absence of any constraints, the master problem will not open any facilities at all. Consequently, the subproblem will return a feasibility cut, because non of the customers can be assigned to a facility. Many more of these iterations may follow until finally a feasible solution is discovered. Many of these iterations could have been avoided if you had added a constraint to the master problem stating that at least a minimum number of facilities should be opened in order to permit a feasible solution (this minimum could be determined by solving a bin-packing problem).