A Chain of Draft (CoD) prompting is a groundbreaking technique that significantly enhances the efficiency of large language models (LLMs) by reducing the amount of text required and processing time. Developed by researchers at Zoom AI in February 2025, this method mimics the natural conciseness of human problem-solving, making it up to 92% cheaper while maintaining or even improving the accuracy of responses.
The Evolution of Prompting Techniques
As we explore the use of large language models, it’s essential to understand the various methods we’ve employed to formulate our requests. Tracing the origins and evolution of this field clarifies the significance of Chain of Draft.
From Basic Prompting to Chain of Thought
In traditional, straightforward prompting, we directly pose a question to the model, which promptly provides an answer without detailed explanation. While this method is token-efficient, it often yields inaccurate results for complex inquiries. For instance, asking, “How much does a business pay after a 27% VAT on a 2 million forint purchase?” may result in a quick but potentially incorrect answer without elaboration.
The next evolution is the Chain of Thought (CoT) prompting, which encourages the model to outline its reasoning step-by-step. Sometimes referred to as “thought chain prompting” in Hungarian literature, this approach is particularly beneficial for more complex tasks as it allows the model to recognize and correct mistakes during the detailed reasoning process.
While the Chain of Thought method significantly improves accuracy, it has a downside: it tends to be quite verbose, using a lot of tokens and slowing down response times. Even a simple computation might come with an extended explanation, increasing costs and wait times.
Chain of Draft: The Power of Brevity
The Chain of Draft prompting retains the benefits of step-by-step reasoning while drastically cutting down on verbosity. The concept is straightforward: we instruct the language model to think step-by-step but to express each thought in roughly five words. This approach mimics how people naturally externalize their thought processes when tackling complex problems—creating minimal outlines that capture only the critical intermediate results without detailing every nuance.
How Does Chain of Draft Prompting Work in Practice?
Implementing Chain of Draft prompting is surprisingly simple. Essentially, we give the model a directive like: “Think step-by-step, but provide only minimal outlines for each thought, using a maximum of 5 words. At the end, provide the final answer after a #### separator.”
Illustrative Examples
Let’s look at a specific example to see how the same task transforms across different prompting techniques:
Example Task: “Zsuzsi invested 120,000 forints, yielding an annual return of 8%. How much money will she have in 3 years?”
Standard Prompt Answer:
“Zsuzsi will have 150,994 forints in 3 years.”
Chain of Thought Prompt Answer:
“Zsuzsi initially invested 120,000 forints, yielding an annual return of 8%. At the end of the first year, her money will be: 120,000 forints × (1 + 0.08) = 120,000 forints × 1.08 = 129,600 forints. At the end of the second year, her money will be: 129,600 forints × 1.08 = 139,968 forints. At the end of the third year, her money will be: 139,968 forints × 1.08 = 151,165.44 forints, or rounded to 151,165 forints. Therefore, Zsuzsi will have 151,165 forints in 3 years.”
Chain of Draft Prompt Answer:
“120000 × 1.08 = 129600. 129600 × 1.08 = 139968. 139968 × 1.08 = 151165.44. #### 151,165 forints.”
It’s clear that Chain of Draft preserves the essential calculation steps while discarding all unnecessary explanatory text, resulting in the same correct outcome with far fewer tokens.
Similarly, it works effectively for logical reasoning tasks:
Example Task: “If all roses are red, and some flowers are roses, what can we conclude for sure?”
Chain of Draft Answer:
“Roses are red. Some flowers are roses. Conclusion: Some flowers are red. #### Some flowers are red.”
Business Benefits and Application Areas
Chain of Draft prompting offers several tangible advantages for business users:
1. Significant Cost Savings
Research indicates that CoD can reduce token usage by 80-90% compared to Chain of Thought prompting. Since most commercial LLM APIs operate on a token-based pricing model (where costs correlate with text volume), this translates to direct savings, especially for high-traffic applications or extensive AI interactions.
2. Faster Response Times
Processing fewer tokens drastically reduces response times. Experiments have shown that Chain of Draft typically cut latency by an average of 76.2% for GPT-4o and 48.4% for Claude 3.5. This is particularly crucial for real-time applications, where quick responses are vital, such as customer service chats, mobile assistants, or interactive tools.
3. Preservation or Improvement of Accuracy
Perhaps most surprisingly, these efficiency gains do not come at the expense of accuracy—in some cases, they even enhance it. Studies show that CoD maintains or improves accuracy (reducing hallucinations) across various reasoning tasks compared to Chain of Thought. For example, in symbolic reasoning tasks, CoD achieved 100% accuracy while reducing token usage by 68-86% (related article).
Practical Application Areas
Chain of Draft can be particularly valuable in the following business domains:
- Data Analysis and Reporting
In data analysis tasks, CoD can help efficiently perform calculations and succinctly articulate conclusions. For example:
Prompt: “Analyze the following quarterly sales data and calculate the growth rate. Use the Chain of Draft method, using a maximum of 5 words for each step, then provide the final conclusion after ####.”
- Document Processing and Information Extraction
When analyzing contracts, reports, or other business documents, CoD can assist in efficiently extracting key information:
Prompt: “Read the following contract and identify the key terms. Use the Chain of Draft method, marking critical points in a maximum of 5 words for each step, then summarize the essence of the contract after ####.”
- Business Decision Support
For complex business decisions, CoD can facilitate a quick and efficient weighing of various factors:
Prompt: “Evaluate the following investment opportunity based on the provided criteria. Using the Chain of Draft method, provide the return calculations in a maximum of 5 words for each step, then give the final recommendation after ####.”
Tips and Best Practices
To effectively apply Chain of Draft, consider the following tips:
1. Clear Instructions
Frame the instructions precisely, for example: “Think step-by-step, but express each step in a maximum of 5 words. Provide the final answer after ####.”
2. Provide Examples
Including one or two examples in the prompt can help illustrate the expected format. This aids the model in understanding the type of concise responses we anticipate.
3. Use Clear Separators
Always utilize a clear separator (e.g., ####) between the thought process and the final answer. This simplifies the extraction of the final result for automated processing.
4. Adjust for Task Type
Slightly modify the instructions for different types of tasks:
- For mathematical problems: “Solve this problem using the Chain of Draft method. Provide only the key calculations in a maximum of 5 words for each step.”
- For logical problems: “Evaluate this logical problem using the Chain of Draft method. Express each reasoning step in a maximum of 5 words.”
5. Iterative Refinement
If the initial response isn’t accurate enough or is too terse, don’t hesitate to refine the prompt. For example, you might ask for a bit more detail in certain steps.
Summary
Chain of Draft prompting can be an exceptionally effective tool for optimizing AI usage. Its straightforward principle—recording the steps of thought with minimal words—can dramatically reduce token usage, response times, and costs while preserving or even enhancing the accuracy of answers.
This innovative technique is especially valuable for companies that engage in high volumes of AI interactions, pay according to text (token) quantity, or require real-time responses. Chain of Draft can be easily implemented with any existing language model and does not require special model modifications or fine-tuning, making it immediately applicable in practice.
