Skeleton Of Thought: Definition and Examples

Skeleton of Thought (SoT) is a prompting method inspired by the human cognitive process: before writing a complex text, one starts by sketching an outline. Applied to large language models, this technique breaks generation into two distinct phases. In the first phase (skeleton stage), the model produces a concise skeleton — a list of key points or a structured outline covering the entire expected response. In the second phase (point-expanding stage), each point of the skeleton is independently developed into a paragraph or full section.

The main benefit of SoT lies in its ability to significantly reduce generation latency. Since each skeleton point can be developed independently, expansions can theoretically be parallelized via simultaneous API calls, reducing response time up to 2x compared to standard sequential generation. This approach was formalized in a research paper published by researchers from Microsoft Research and Tsinghua University in 2023.

Beyond speed, SoT often improves response quality. By forcing the model to plan before writing, it avoids digressions, repetitions, and structural inconsistencies common in long responses. The skeleton acts as a mental map guiding generation and ensuring comprehensive coverage of the topic.

It is important to note that SoT is not suitable for all tasks. Simple questions, coding tasks, or queries requiring strict sequential reasoning (like step-by-step math problems) benefit little from this approach. SoT excels instead for long, structured, multi-faceted responses such as analyses, comparisons, or explanatory guides.