We explore the evolution of programs for classification tasks, using the recently introduced Hierarchical Evolutionary Re-Combination Language (HERCL) which has been designed as an austere and general-purpose language, with a view toward modular evolutionary computation, combining elements from Linear GP with stack-based operations from forth. We show that evolved HERCL programs can successfully learn to perform a variety of benchmark classification tasks, and that performance is enhanced by the sharing of genetic material between tasks.