A prompt is input that gives a model instructions, examples, or context. A well-crafted prompt can improve performance without changing the model, but it helps reveal existing capabilities rather than creating new ones. section 7.5.1 section 7.5.3
A familiar puzzle with changed rules
In the Wolf, Goat, and Cabbage puzzle, a farmer ferries the animals and cabbage across a river under safety constraints. A variant changes which animals eat which; a model pulled toward the familiar solution may overlook the new constraints. The prompt can shift its response, but the failure reflects how strongly its answer follows familiar patterns. section 7.5.1
Prompts can make reasoning more explicit. Representing each state as a triple, listing legal moves, and checking safety after each crossing can help catch invalid steps. Chain-of-thought prompting similarly asks for intermediate reasoning before the final answer; it has improved performance on multistep math and commonsense tasks without additional training. section 7.5.3 section 7.5.3
