Chain of Draft (CoD) Prompting Method

A Chain of Draft (CoD) Prompting Method is a prompt engineering method that encourages large language models to generate minimalistic yet informative intermediate reasoning outputs while solving tasks.

• AKA: CoD Prompting, Minimalistic Reasoning Prompting, Concise Thought Prompting.
• Context:
  • It can typically reduce token usage through minimalistic intermediate steps while maintaining or improving task accuracy.
  • It can typically focus on essential calculations and critical insights rather than verbose explanations.
  • It can typically achieve similar reasoning effectiveness as chain-of-thought prompting with significantly fewer output tokens.
  • It can typically decrease inference latency by reducing the computational resources required for generating responses.
  • It can typically lower computational costs by minimizing both input tokens and output tokens during reasoning processes.
  • ...
  • It can often be implemented through simple prompt instructions that encourage brevity in each reasoning step.
  • It can often use instructions like "keep a minimum draft for each thinking step, with 5 words at most" to guide model behavior.
  • It can often improve real-time application performance where low latency is crucial.
  • It can often preserve reasoning transparency while eliminating redundant verbosity.
  • ...
  • It can range from being a Zero-Shot CoD Prompting Method to being a Few-Shot CoD Prompting Method, depending on its example provision.
  • It can range from being a Simple Mathematical CoD Method to being a Complex Symbolic CoD Method, depending on its task domain.
  • It can range from being a Token-Sensitive CoD Method to being a Per-Step Budgeted CoD Method, depending on its constraint approach.
  • ...
  • It can have human-inspired cognitive approaches that mimic how people jot down essential information when solving problems.
  • It can have per-step budget allocation rather than global token limitations, allowing unlimited reasoning steps while constraining each individual step.
  • It can have adaptive token usage that proportionally matches the task complexity.
  • ...
  • It can be complementary to other latency-reducing techniques such as draft-and-verify methods and skeleton-of-thought approaches.
  • It can be particularly effective for arithmetic reasoning tasks, commonsense reasoning tasks, and symbolic reasoning tasks.
  • It can be more efficient than chain-of-thought prompting while achieving comparable or better performance metrics.
  • ...
• Examples:
  • CoD Prompting Implementations, such as:
    • Mathematical CoD Prompts, such as:
      • Arithmetic Problem CoD for mathematical calculations, where expressions like "20 - x = 12; x = 8" replace verbose explanations.
      • Algebraic CoD Sequence for equation solving, where each step contains only the essential mathematical transformation.
    • Symbolic CoD Prompts, such as:
      • Coin Flip CoD for state tracking problems, using minimal notation to track state changes.
      • Logic Puzzle CoD for constraint satisfaction problems, using shorthand to track valid possibilities.
  • CoD Prompt Instructions, such as:
    • Minimalist Guidelines CoD that instruct "Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most."
    • Equation-Only CoD that specifies "Show only equations and essential variables, no explanations."
  • ...
• Counter-Examples:
  • Chain-of-Thought (CoT) Prompting Method, which encourages verbose step-by-step reasoning with detailed explanations rather than minimalistic drafts.
  • Verbose Reasoning Method, which intentionally maximizes explanation detail for educational or transparency purposes.
  • Standard Direct Prompting, which provides only final answers without any intermediate reasoning steps or drafts.
  • Skeleton-of-Thought Method, which focuses on creating an outline first and then expanding it, rather than maintaining minimal drafts throughout.
• See: Token Optimization, Reasoning Efficiency, Inference Latency Reduction, Human-Inspired Cognition Model, Minimalistic Reasoning.

---

---

References

2025

• (Xu et al., 2025) ⇒ Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. (2025). “Chain of Draft: Thinking Faster by Writing Less.” doi:10.48550/arXiv.2502.18600

---