TL;DR
A prompting technique that instructs language models to generate concise, highly compressed intermediate drafts during reasoning to save tokens.
This methodology optimizes standard reasoning prompting by constraining the output of intermediate logical steps to short, highly dense phrases rather than verbose prose. It instructs the language model to write its thinking steps as minimal drafts, often capping each step at a strict limit of five words. By doing so, it preserves the logical benefits of step-by-step reasoning while reducing reasoning token usage by up to eighty percent, significantly lowering latency and API costs.
Why this matters for your business
It allows enterprises to deploy high-quality reasoning capabilities in production environments at a fraction of the traditional cost and latency. This makes real-time, step-by-step AI workflows much more commercially viable.
