Process discovery aims to discover models that can explain the behaviors of event logs extracted from information systems. While various approaches have been proposed, only a few guarantee desirable properties such as soundness and free-choice. State-of-the-art approaches that exploit the representational bias of process trees to provide the guarantees are constrained to be block-structured.Such constructs limit the expressive power of the discovered models, i.e., only a subset of sound free-choice workflow nets can be discovered. To support a more flexible structural representation, we aim to discover process models that provide the same guarantees but also allow for non-block structures. Inspired by existing works that utilize synthesis rules from the free-choice nets theory, we propose an automatic approach that incrementally adds activities to an existing process model with predefined patterns. Playing by the rules ensures that the resulting models are always sound and free-choice. Furthermore, the discovered models are not restricted to block structures and are thus more flexible. The approach has been implemented in Python and tested using various real-life event logs. The experiments show that our approach can indeed discover models with competitive quality and more flexible structures compared to the existing approach.