Over the past decade, high contrast imaging has afforded astronomers the opportunity to isolate the light directly emitted by dozens of newly-discovered worlds. The exoplanet direct imaging technique promises ultra-detailed characterization of planetary atmospheres; however, the immense brightness of starlight in extrasolar systems can mask dim planetary signals, perpetually limiting the technique. Fortunately, recent software developments have transformed direct imaging by allowing us to empirically model and robustly remove excess starlight. pyKLIP is one of the most modern python libraries with such capabilities. By comparing images taken in a range of on-sky orientations, pyKLIP is able to identify common patterns and use them to model the stellar PSF image-by-image. Unfortunately, the characteristics of the final PSF-subtracted image are greatly dependent on tunable user inputs. For the same dataset, a poor choice of input parameters can completely eliminate robust planet signals, while other combinations of parameters can lead to an unquestionable detection. The lack of a procedure for optimizing KLIP parameters to yield clear detections while avoiding false positives results in an ad hoc by eye optimization approach. Using data collected with the Magellan Telescope’s Adaptive Optics Instrument (MagAO) as part of the Giant Accreting Protoplanet Survey (GAPlanetS), we explore the entirety of pyKLIP parameter space for several datasets with and without known companions. We report progress on developing a dependable metric for KLIP parameter-tuning to improve companion detections. Upon completion, our parameter optimization techniques will be extended beyond the GAPlanetS survey and exploited to advance observing strategies for future direct imaging surveys and missions. Specifically, we hope to test our optimization method on a subset of simulated 3-5 micron James Webb Space Telescope Near Infrared Camera (NIRCAM) data, where we expect many exoplanet spectral features to be their brightest. This will ultimately allow us to test the robustness of our optimization technique for a range of instruments and observing modes.