Skeleton-of-Thought

Skeleton-of-Thought (SoT) is a prompting pattern that first asks the model to produce a short "skeleton" — a bare list of the points its answer should cover — and then expands each point separately. Because the expansions don't depend on one another, they can be generated in parallel rather than one token after another, which is the main draw: SoT (Ning et al., 2023) was introduced primarily to cut the latency of long answers, sometimes by roughly 2x, while often keeping or improving quality on list-like tasks. For SaaS builders, the practical angle is response speed on features that produce structured, multi-part output — think "generate 5 onboarding email drafts" or a comparison breakdown — where users feel every second of streaming delay. The trade-off is orchestration complexity: you make one call for the outline, fan out parallel calls for each section, then stitch them together. It suits enumerable answers, not tightly reasoned chains where each step depends on the last.