We describe an approach to tracing the “step-by-step” computation involved when a model responds to a single prompt.