This talk presents a method for automating the repetitive and mechanical aspects of a compiler engineer's work by encoding optimization plans into an MLIR-based IR. Building on MLIR's Transform dialect, this approach introduces operations with **non-deterministic choice semantics**, allowing engineers to define alternative optimization sub-schedules and parameter choices directly within the transform IR. A driver then uses a solver to explore these choices, rewrite the schedule into a deterministic one, apply it, and feed performance results back to the solver, creating an automated optimization loop. This method goes beyond traditional auto-tuning by manipulating the optimization schedule itself, yet differs from auto-scheduling by allowing the engineer to provide a template with specific choice points, thereby maintaining control while achieving powerful automation.