Hybrid RANS/LES methods can produce more reliable results than RANS with a reasonable computational cost. Thus, they have the potential to become the next workhorse in the industry. However, in continuous approaches, the location of the switching between the RANS and LES modes is based on the mesh generated by the user, such that the results are user-dependent. The present paper aims at developing a self-adaptive strategy, in which the RANS and LES zones are determined based on physical criteria, in order to mitigate the influence of the user. Starting from an initial RANS computation, successive HTLES are carried out and the mesh is refined according to the criteria, which are discussed at the beginning of the paper. In order to demonstrate the feasibility of this strategy, the method is applied to the case of a backward-facing step with the Hybrid Temporal Large Eddy Simulation (HTLES) approach, but is suitable for any other hybrid approach. The results obtained show that the method converges after only a few simulations and significantly improves the predictions when compared to RANS, with no intervention from the user. The power spectral density plots and Q-criterion visualization highlight the physical consistency of the results and the comparison of statistically averaged quantities with the DNS is very encouraging. Even though the process is still a long way from being applicable to a wide range of turbulent flows, this paper is a demonstrator of the applicability of this self-adaptive strategy.