What can humans write in prompts that the model cannot do by itself?

This question defines the boundary of what needs to be built into agent runtimes versus what can be left to the model. It emerged from observing that some prompt constructs genuinely expand model capabilities while others are redundant.