MLIR dialect design is often more an art than a science. MLIR provides powerful infrastructure for building IR and a vast ecosystem of dialects to use, but lacks guidelines on how to actually do so. This talk is a deep dive on principles for MLIR dialect design and composition. We will focus on criteria for dialect design, such as concise and powerful representation and transparent composability, dialect design principles and techniques, such as preserving high-level information and distinguishing between “structural” and “computation” dialects, and challenges with integrating with upstream dialects. We will study how good dialect design allows us efficiently and easily write powerful optimizations on our IR and how to write generic IR transformations with MLIR interfaces. We will present our findings through the lens of building a general-purpose programming language with MLIR.