Skip to content
Back to full roadmap
topicadvanced★ Pro

Skeleton-of-Thought

First produce the skeleton (outline), then fill sections IN PARALLEL — 2-3× latency reduction.

2 hours1 resources1 prereqs

Classic CoT is sequential — long answers are slow. SoT is two-stage:

  1. Tell the model to "produce a 5-point skeleton"
  2. Make a parallel API call per point

Great for blog posts, reports, stories. Trade-off: cost 1.x → 5× per point, but wall-clock drops ~50%.

Prerequisites

Resources(1)