Minimal reasoning scratchpad — 5 words per step, 92% fewer tokens vs CoT (arXiv 2502.18600)
Chain of Draft (CoD) Prompting Technique Source: "Chain of Draft: Thinking Faster by Writing Less" — arXiv 2502.18600 (Feb 2025) ------------------------------------------------------------------ TECHNIQUE OVERVIEW: Chain of Draft constrains each reasoning step to ≤5 words, forcing minimal but essential intermediate thinking. On GSM8k math: 91% accuracy vs CoT's 95%, using only 7.6% of tokens. Latency reduction up to 76%. ------------------------------------------------------------------ BASIC SYSTEM PROMPT: Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most. Return the answer after a #### separator. ------------------------------------------------------------------ EXTENDED SYSTEM PROMPT (for complex tasks): Think step by step. For each reasoning step, write a minimal draft of 5 words or fewer — only the essential operation or transformation. Do not explain; just note what you are doing. After all steps, write #### on its own line, then give the final answer. Example format: Step 1: [≤5 words] Step 2: [≤5 words] Step 3: [≤5 words] #### [Final answer here] ------------------------------------------------------------------ USAGE NOTES: - Best for: math word problems, logical reasoning, multi-step calculations - Not ideal for: creative writing, open-ended generation, tasks requiring explanation - Works with: GPT-4, Claude 3+, Gemini 1.5+ - Token savings: ~92% vs standard CoT; latency: up to 76% lower - Accuracy tradeoff: ~4% below CoT on math benchmarks — acceptable for most real-world use ------------------------------------------------------------------ FEW-SHOT EXAMPLE: User: "Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 balls. How many tennis balls does he have now?" Model: Step 1: 2 cans × 3 balls Step 2: 6 new balls Step 3: 5 + 6 = 11 #### 11
