Back to full roadmap
topicadvanced
Skeleton-of-Thought Planning
Skeleton (outline) first, then fill in parallel — 2-3× latency reduction.
2 hours1 resources1 prereqs
For long answers (blog post, report) sequential CoT is slow. Skeleton-of-Thought:
- "Produce a 5-point skeleton"
- Parallel API call per point
- Merge results
Trade-off: cost × number of points, but wall-clock drops 50-70%. Big UX win.