RIP Chain of Thought. I just tested Skeleton of Thought prompting and it's cutting response latency in half on structured tasks while improving output quality. One paper changed how I think about AI text generation. CoT forces the model to think and write simultaneously. Point 1 must finish before Point 2 starts. Point 2 before Point 3. Every token blocked by the token before it. Experts don't work that way. A McKinsey consultant outlines every section first, then develops each one independently. A professor skeletons the topics, then fills them in. SoT matches that. Phase 1 - Skeleton: The model writes a bare outline of every point. No elaboration. Just structure. Phase 2 - Parallel Fill: Each skeleton point expands independently. Point 3 doesn't wait for Point 2. One wrong branch doesn't slow down every branch after it. The exact prompt: "First, write a concise skeleton of your response as a list of key points. Then, expand each point independently and in full. Format: [Skeleton] followed by [Point 1], [Point 2], etc." Where SoT dominates: → How-to guides and tutorials → Listicles and ranked breakdowns...