There has been significant research interest in Procedural Level Generation via Machine Learning (PLGML), applying ML techniques to automated level generation. One recent trend is in the direction of learning representations for level design via embeddings, such as tile embeddings. Tile Embeddings are continuous vector representations of game levels unifying their visual, contextual and behavioural information. However, the original tile embedding struggled to generate levels with skewed tile distributions. For instance, Super Mario Bros. (SMB) wherein a majority of tiles represent the background. To remedy this, we present a modified tile embedding representation referred to as Clustering-based Tile Embedding (CTE). Further, we employ clustering to discretize the continuous CTE representation and present a novel two-step level generation to leverage both these representations. We evaluate the performance of our approach in generating levels for seen and unseen games with skewed tile distributions and outperform the original tile embeddings.