The results of over twenty-five years of research seem clear: the addition of seductive visual details in video games hinders performance of learners (Garner, Gillingham, & White, 1989; Thalheimer, 2004; Rey, 2012). Yet, countless other research results propose the opposite: that visual embellishments and well-designed ambiguity instead "improve" learners' performance, engagement, and self-efficacy (Tierney, Corwin, Fullerton, & Ragusa, 2014; Wilson et. al 2009; Scott & Ghinea, 2013). To shed light on this apparent contradiction, we devised a particular experiment using "game skins" to implement variations in visual themes of a computer game. Game skins are coherent, interchangeable sets of graphical assets that all implement the same underlying game structure while varying the visual appearance. In particular, we implemented the following four game skins labeled and described as follows: (1) "Generic" theme with no embellishments (simple flat color background), (2) "Fantasy" game theme (forest, snow, and desert adventure backgrounds), (3) "STEM-oriented" theme (computer circuitry background), and (4) "Choice" (the user picks one of the previous three options). Our goal is determining if there are differences in performance, engagement, and self-efficacy between conditions. The upshot is that the generic condition participants had highest performance (levels) and had highest programming self-efficacy--followed by choice, fantasy game setting, circuitry. However, ordering of conditions for engagement was precisely opposite the trend for performance. We conclude by discussing the trade-offs between the two diametrically opposed approaches to game themes and embellishment: instrumental game skins vs. thematic and deliberately embellished game skins.