What is Chain of Draft?
Chain of Draft (CoD) is a prompting technique from the 2025 paper Chain of Draft: Thinking Faster by Writing Less by Silei Xu and colleagues at Zoom Communications. It starts from an observation about how people solve problems: we rarely write out every thought in full sentences. We jot down the essentials, such as a number, an equation or a keyword, and move on.
Chain-of-Thought prompting makes models do the opposite. It produces long, verbose explanations, which cost tokens and time. Chain of Draft keeps the step-by-step structure but limits each step to a minimal draft, so the model still reasons explicitly while writing far less.
The instruction used in the paper is short:
Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most. Return the answer at the end of the response after a separator ####.
On the GSM8K math benchmark, GPT-4o with Chain-of-Thought used about 205 tokens per answer at 95.4% accuracy. With Chain of Draft it used about 44 tokens at 91.1% accuracy, and latency fell from about 4.2 to 1.0 seconds. Claude 3.5 Sonnet showed a similar pattern. On date understanding and sports understanding tasks, Chain of Draft matched or exceeded Chain-of-Thought accuracy with a fraction of the tokens.
When to Use Chain of Draft
High-Volume API Workloads
Classification, extraction and scoring pipelines where reasoning helps accuracy but verbose output multiplies your bill.
Latency-Sensitive Apps
Chat assistants, autocomplete and voice interfaces where users notice every extra second of generated reasoning.
Arithmetic and Logic
Word problems and calculations where the essential steps are numbers and short equations, not prose.
Cheap Self-Consistency
Sampling several reasoning paths becomes affordable when each path is a few dozen tokens instead of a few hundred.
Context-Limited Settings
Long agent loops or small context windows where verbose reasoning would crowd out the information that matters.
Structured Decisions
Rule checks, eligibility decisions and routing where each step is a condition and a result.
How to Use Chain of Draft
- 1
Add the draft instruction
Tell the model to think step by step but keep each step to a minimal draft of about five words, and to put the final answer after a clear separator such as ####.
- 2
Include two or three draft-style examples
The paper found zero-shot CoD much weaker. Show examples whose reasoning is written as terse drafts (for example "20 - x = 12; x = 8") so the model copies the format.
- 3
Parse the answer after the separator
Because the reasoning is compact, extracting the answer is easy: split on the separator and keep what follows.
- 4
Benchmark against Chain-of-Thought
Run a small test set with both prompts. Keep CoD where accuracy holds and fall back to full CoT for the task types where it drops.
Prompt Examples
Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most. Return the answer after ####. Q: Jason had 20 lollipops. He gave Denny some. Now he has 12. How many did he give Denny? A: 20 - x = 12; x = 8. #### 8 Q: A shop sells pens at 3 for $4. How much do 12 pens cost? A: 12 / 3 = 4 packs; 4 * 4 = 16. #### 16 Q: A train travels 180 km in 2.5 hours. What is its average speed in km/h? A:
Decide whether this refund request is eligible. Think step by step, but keep each step to 5 words at most. Return ELIGIBLE or NOT ELIGIBLE after ####. Policy: Refunds within 30 days of delivery, item unused, not a final-sale item. Example Request: Delivered 12 days ago, unopened, regular item. Draft: 12 days < 30; unused yes; not final sale. #### ELIGIBLE Request: Delivered 41 days ago, unopened, regular item. Draft:
Pros and Cons
| 🟢 Pros | 🔴 Cons |
|---|---|
| Roughly 70–90% fewer reasoning tokens than Chain-of-Thought in the paper's benchmarks | Small accuracy drop on some tasks (about 4 points on GSM8K) |
| Much lower latency, which users feel directly | Weak zero-shot: needs draft-style examples |
| Keeps an explicit, checkable reasoning trace | Terse drafts are harder for humans to audit than full explanations |
| Makes Self-Consistency and agent loops cheaper | Larger gap to CoT on small models (under ~3B parameters) |
Chain of Draft Prompt Template
Copy this template and replace each {{ variable }}. It uses Prompt Edit's variable syntax, so you can paste it into the app and fill the fields from a form every time you reuse it.
Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most. Return the answer at the end after ####.
Q: {{ example_question_1 }}
A: {{ example_draft_1 }} #### {{ example_answer_1 }}
Q: {{ example_question_2 }}
A: {{ example_draft_2 }} #### {{ example_answer_2 }}
Q: {{ question }}
A:Frequently Asked Questions
What is Chain of Draft prompting?
Chain of Draft (CoD) is a reasoning technique introduced by Xu et al. at Zoom Communications in 2025. Like Chain-of-Thought, it asks the model to reason step by step, but each step must be a minimal draft of about five words. The model keeps only the essential numbers, equations or facts and returns the final answer after a separator. The result is far shorter reasoning at similar accuracy on many tasks.
What is the Chain of Draft prompt?
The prompt used in the paper is: 'Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most. Return the answer at the end of the response after a separator ####.' The paper pairs it with a few worked examples written in the same terse draft style.
How many tokens does Chain of Draft save?
In the paper, Chain of Draft used as little as 7.6% of the tokens of Chain-of-Thought on some tasks. On the GSM8K math benchmark it cut reasoning tokens by roughly 80% (for example from about 205 to 44 tokens on GPT-4o) and latency by about 75%, while accuracy dropped from about 95% to about 91%. On other tasks, such as date and sports understanding, it matched or beat Chain-of-Thought.
Is Chain of Draft better than Chain-of-Thought?
It is cheaper and faster, not universally better. Use Chain of Draft when latency and token cost matter and a small accuracy trade-off is acceptable. Keep full Chain-of-Thought when the reasoning itself must be readable (for audits, teaching or debugging) or when every percentage point of accuracy counts.
Does Chain of Draft work zero-shot?
Not reliably. The authors report a significant drop in the zero-shot setting and conclude that the terse draft style is rare in training data, so the model needs a few examples written in draft style. Always include two or three short examples.
Does Chain of Draft work with small models?
Less well. The gap to Chain-of-Thought was more pronounced for models under about 3 billion parameters. Chain of Draft is best suited to capable frontier models.
Should I use Chain of Draft with reasoning models?
Reasoning models already think internally before answering, so an explicit visible draft adds little. Chain of Draft is most useful with standard instruction-tuned models where you pay for every visible reasoning token.
Origin & Sources
- Xu et al. (Zoom Communications) (2025). Chain of Draft: Thinking Faster by Writing Less. arXiv.
- Wei et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv / NeurIPS 2022.