least-to-most prompting

Least-to-most prompting tackles a hard problem by first breaking it into a ladder of smaller subproblems, then solving them in order from easiest to hardest, so that each answer becomes a stepping stone for the next. Rather than demanding the full solution in one go, you ask the model what simpler questions it would need to answer first, and then walk up that list. The earlier solved pieces stay in the context, giving the later steps the parts they need.

It helps most when a problem is too tangled to solve directly but naturally splits into stages, like a multi-part word problem where you must find one quantity before you can compute the next. By making the decomposition explicit, you keep the model from skipping the groundwork and leaping to a shaky answer.

It is close cousin to chain-of-thought, but with a distinction: chain-of-thought reasons step by step within one answer, while least-to-most first plans the subproblems and then solves them, sometimes across separate model calls. The risk is that a bad decomposition leads everything astray, so the quality of the breakdown matters as much as the solving.