Once it’s in the context window the model invariably steers crazy. Llms cannot handle the “don’t think of an elephant” requirement.