Constraint satisfaction problems occur widely in artificial intelligence. Hence, arc consistency techniques have been widely studied to simplify constraint networks before or during the search for solutions. To reduce the cost of maintenance, the researchers have focused their work on the improvement of maintaining a single arc consistency. In this paper, from a higher point of view, the authors try to propose some principles and the corresponding strategies of three levels, which are search level, maintenance level and arc level. In this way, MAC-Dynamic and AC-I⁺ are presented. The effectiveness of this approach is demonstrated experimentally on two typical benchmarks of CSPs: Zebra Puzzles and N-Queen Problem.