prompt logo

prompt

Free

Think faster by writing less — minimal reasoning for efficient LLM prompting.

FreeFree tier
Inputs: textOutputs: text
Type
Open Source

About prompt

Chain of Draft (CoD) is a prompting technique that constrains each reasoning step to a maximum of 5 words, forcing minimal but essential intermediate thinking. It significantly reduces token usage and latency compared to traditional Chain of Thought prompting, while maintaining high accuracy on math and logical reasoning tasks. According to the paper (arXiv 2502.18600), CoD achieves 91% accuracy on GSM8k math benchmarks using only 7.6% of tokens, with latency reduction up to 76%. The technique is best suited for math word problems, logical reasoning, and multi-step calculations, and works with major LLMs such as GPT-4, Claude 3+, and Gemini 1.5+. System prompts and a few-shot example are provided for easy implementation.

Key Features

Each reasoning step limited to a maximum of 5 words
Up to 92% token savings and 76% lower latency compared to standard Chain of Thought
Maintains 91% accuracy on GSM8k math benchmarks
Simple system prompts provided for easy integration
Compatible with GPT-4, Claude 3+, and Gemini 1.5+

Pros & Cons

Pros
  • Dramatic reduction in token usage and latency compared to Chain of Thought
  • Simple to implement with provided system prompts and few-shot examples
  • Maintains competitive accuracy while being significantly more efficient
  • Works with leading language models without special configuration
Cons
  • Slightly lower accuracy than Chain of Thought (about 4% drop on math benchmarks)
  • Not suitable for creative writing, open-ended generation, or tasks requiring detailed explanation
  • 5-word constraint may limit expressiveness for very complex reasoning steps

Best For

Math word problemsLogical reasoningMulti-step calculations

FAQ

What is Chain of Draft (CoD) prompting?
Chain of Draft is a prompting technique that constrains each reasoning step to a maximum of 5 words, forcing minimal but essential intermediate thinking. It aims to reduce token usage and latency while maintaining accuracy on tasks like math word problems and logical reasoning.
How does CoD compare to Chain of Thought (CoT)?
CoD uses far fewer tokens (about 7.6% of CoT's tokens) and reduces latency by up to 76%, with only a minor accuracy drop (e.g., 91% vs 95% on GSM8k math). It trades a small amount of accuracy for significant efficiency gains.
Which models are compatible with Chain of Draft?
The technique works with GPT-4, Claude 3+, and Gemini 1.5+, as stated in the usage notes.
What types of tasks is CoD best suited for?
CoD is best for math word problems, logical reasoning, and multi-step calculations. It is not ideal for creative writing, open-ended generation, or tasks requiring detailed explanation.