I briefly review the state-of-the-art spatial regression discontinuity (RDD) framework that is suitable to isolate causal effects when there are sharp spatial breaks. I then propose several improvements. First, I illustrate that some commonly used specifications are prone to type-I errors, especially when there are spatial trends in the data. Second, I propose a way to report heterogeneous treatment effects alongside the RD cutoff. Third, I introduce randomization inference to the spatial RD framework by creating a set of functions that allow to randomly shift borders. These tools might be interesting for other identification strategies that rely on the shift of boundaries. A companion R-package called SpatialRDD includes all the tools necessary to carry out spatial RD estimation, including the proposed improvements. Furthermore, I address recent concerns that spatial RDDs might suffer from spurious regression problems due to spatially correlated omitted variables - or spatial autocorrelation more generally. Using spatial Monte Carlo simulations I demonstrate the opposite. In the presence of such phenomena, spatial RDDs are actually a solution when it comes to too small standard errors in frequentist hypothesis testing induced by spatial correlation.