The archaeological record represents a window onto the complex relationship between stone artefact variance and hominin behaviour. Differences in the shapes and sizes of stone flakes—the most abundant remains of past behaviours for much of human evolutionary history—may be underpinned by variation in a range of different environmental and behavioural factors. Controlled flake production experiments have drawn inferences between flake platform preparation behaviours, which have thus far been approximated by linear measurements, and different aspects of overall stone flake variability (Dibble and Rezek J Archaeol Sci 36:1945–1954, 2009; Lin et al. Am Antiq 724–745, 2013; Magnani et al. J Archaeol Sci 46:37–49, 2014; Rezek et al. J Archaeol Sci 38:1346–1359, 2011). However, when the results are applied to archaeological assemblages, there remains a substantial amount of unexplained variability. It is unclear whether this disparity between explanatory models and archaeological data is a result of measurement error on certain key variables, whether traditional analyses are somehow a general limiting factor, or whether there are additional flake shape and size drivers that remain unaccounted for. To try and circumvent these issues, here, we describe a shape analysis approach to assessing stone flake variability including a newly developed three-dimensional geometric morphometric method (‘3DGM’). We use 3DGM to demonstrate that a relationship between platform and flake body governs flake shape and size variability. Contingently, we show that by using this 3DGM approach, we can use flake platform attributes to both (1) make fairly accurate stone flake size predictions and (2) make relatively detailed predictions of stone flake shape. Whether conscious or instinctive, an understanding of this geometric relationship would have been critical to past knappers effectively controlling the production of desired stone flakes. However, despite being able to holistically and accurately incorporate three-dimensional flake variance into our analyses, the behavioural drivers of this variance remain elusive.