Scientific workflow often requires dynamic selection of workflow routines, Web Services or workflow engines. Multiple copies of a Web service or multiple workflow engines with different performance are chosen at run time to optimise the workflow. However, simple performance formula for selecting Web services or workflow engines is difficult to find. In some cases, the Services with the same functions but with different algorithms (for example, data clustering services implemented by using the algorithms of Neural Network or SVM) may be only chosen at the just pre-execution point according to the intermediate results of workflow execution. We here use production rule strategy to solve the above problem. In this paper, we present a framework for production rule based selection decision for dynamic workflow, and give out a primary implementation of this framework.