Efficient Reasoning

Chain of Draft

Show your work, but in shorthand. Chain of Draft keeps the step-by-step reasoning that makes Chain-of-Thought reliable, then strips each step down to a minimal draft of just the load-bearing words. The model reaches the same answer while writing a small fraction of the tokens, which means lower cost and faster responses.

*These are learning tools, not prompts. They teach you to write your own, think of them as training wheels that guide you while you learn, then fall away once you can ride on your own.

The Core Insight

Reasoning in Shorthand

When a person works through a hard problem on scratch paper, they rarely write full sentences. They jot the equation, the intermediate number, the key fact, just enough to hold the thread. The reasoning is real, but the prose around it is not. Chain of Draft asks a language model to reason the same way: keep every thinking step, but compress each one to a terse draft instead of a fluent paragraph.

The insight is that fluency is not reasoning. Standard Chain-of-Thought produces long, well-written explanations, and most of those words are connective tissue, not the actual logical move. Chain of Draft keeps the logical moves and discards the connective tissue. In the original study, the model was instructed to keep each step to a minimal draft of only a handful of words, and it reached comparable accuracy while using as little as 7.6 percent of the tokens that full Chain-of-Thought required.

Fewer tokens is not just a tidiness win. Output tokens drive both the price and the latency of a response, so a reasoning trace that is one twelfth the length is roughly one twelfth the cost and lands much faster. For high-volume or latency-sensitive work, that difference is the difference between affordable and not.

Why Terse Drafts Keep the Accuracy

The accuracy of step-by-step reasoning comes from the model committing to intermediate results before it answers, not from how verbosely it explains them. A draft like "20 - 8 = 12" commits to the same intermediate value as a full sentence describing the subtraction, and the model can condition on it just as well. Chain of Draft preserves the part that matters, the explicit intermediate step, and removes the part that does not, the narration.

The Chain of Draft Process

Four stages from question to answer, with the reasoning kept deliberately terse

1

Pose the Problem with a Draft Constraint

Give the model the question along with an instruction to think step by step but to keep each step to a minimal draft of only a few words. This single constraint is what separates Chain of Draft from ordinary Chain-of-Thought; it tells the model to record its reasoning as notes, not prose.

Example

"Jason had 20 lollipops. He gave some to Denny. Now he has 12. How many did he give away? Think step by step, but keep each step to a minimal draft of about five words."

2

Draft the First Reasoning Step

The model writes its first intermediate step as a compact draft, capturing the load-bearing value or fact and nothing else. The goal is to commit to a concrete intermediate result, the same commitment full Chain-of-Thought makes, but without the surrounding explanation.

Example

"Start: 20 lollipops."

3

Continue Drafting to the Conclusion

Each subsequent step is another terse draft that builds on the last. The chain stays short because every line carries only the essential operation or result. The model is reasoning just as deliberately as it would in a long trace, it is simply not narrating.

Example

"Now has 12. 20 - 12 = 8."

4

Return the Final Answer

After the minimal drafts, the model states the final answer clearly, typically after a simple separator so the answer is easy to extract. The user receives a short, scannable trace plus a clean answer, having paid for a fraction of the tokens a verbose chain would have cost.

Example

"#### He gave away 8 lollipops."

See the Difference

Same answer, same reasoning, a fraction of the tokens

Chain-of-Thought

Verbose Reasoning

"Let's work through this carefully. Jason started with 20 lollipops. After giving some away to Denny, he now has 12 lollipops left. To find how many he gave away, we subtract the lollipops he has now from the lollipops he started with. So that is 20 minus 12, which equals 8. Therefore Jason gave away 8 lollipops."

Cost

Correct answer, but every word is billed and adds latency. Most of the tokens are narration that did not change the result.

Accurate, but long, slower, and more expensive
VS

Chain of Draft

Minimal Drafts

"Start: 20.
Now: 12.
20 - 12 = 8.
#### 8"

Cost

Same intermediate steps, same answer, written in a handful of tokens. In the original study this style reached comparable accuracy at as little as 7.6 percent of the tokens.

Same reasoning, far fewer tokens, lower cost and latency

Practice Responsible AI

Always verify AI-generated content before use. AI systems can produce confident but incorrect responses. When using AI professionally, transparent disclosure is both best practice and increasingly a legal requirement.

Most US states are actively legislating AI transparency and accountability. Critical thinking remains your strongest tool against misinformation.

Chain of Draft in Action

See how terse drafts hold up across arithmetic, logic, and everyday reasoning

Prompt

"A cafe sells 18 muffins in the morning and twice as many in the afternoon. If each muffin costs 3 dollars, how much did the cafe earn? Think step by step, but keep each step to a minimal draft of about five words. Put the final answer after ####."

Chain of Draft Response

"Morning: 18 muffins.
Afternoon: 2 x 18 = 36.
Total: 18 + 36 = 54.
Earnings: 54 x 3 = 162.
#### 162 dollars."

Prompt

"Three friends finish a race. Ana is not last. Beto is ahead of Ana. Who finished last? Think step by step in minimal drafts of a few words each, then give the answer after ####."

Chain of Draft Response

"Beto before Ana.
Ana not last.
So Carla last.
#### Carla finished last."

Prompt

"A recipe needs 750 milliliters of stock. You only have a 1-cup measure, where 1 cup is 240 milliliters. How many full cups, and how much is left over? Reason in minimal drafts, answer after ####."

Chain of Draft Response

"750 / 240 = 3.125.
3 full cups = 720.
750 - 720 = 30.
#### 3 cups, 30 ml left."

When to Use Chain of Draft

Best where reasoning matters but verbosity is wasted cost

Perfect For

High-Volume Reasoning Tasks

When you run the same kind of multi-step problem thousands of times, cutting the reasoning trace to a fraction of its length turns a large token bill into a small one.

Latency-Sensitive Interfaces

Shorter output means a faster response. For chat, autocomplete, or any interactive surface where the user is waiting, terse drafts keep reasoning without the wait.

Arithmetic and Symbolic Steps

Tasks whose reasoning is naturally a sequence of concrete values or operations compress cleanly, the draft is just the equation, with nothing lost.

Agent Loops and Tool Chains

Inside an agent that reasons on every turn, shorter intermediate thoughts keep the context window lean and each step cheap, which compounds over a long run.

Skip It When

You Need a Full Audit Trail

When the reasoning itself is the deliverable, a legal memo, a medical rationale, a teaching explanation, the narration is the value. Use full Chain-of-Thought instead.

Very Small Models

The original study found the gains are strongest on capable models. Smaller models can lean on the extra scaffolding of verbose reasoning, so the terse style may cost some accuracy there.

Open-Ended or Creative Work

Brainstorming, drafting, and exploratory writing benefit from the model thinking out loud. Compressing the thinking removes the very thing that makes the output rich.

Use Cases

Where Chain of Draft delivers the most value

Cost-Controlled APIs

Cut the output-token bill on reasoning-heavy endpoints by an order of magnitude while keeping answers correct, turning an expensive feature into an affordable one.

Real-Time Assistants

Keep step-by-step accuracy in chat and voice assistants without the lag of a long reasoning trace, so the answer arrives while the user is still paying attention.

Batch Data Processing

When classifying, extracting, or computing over large datasets row by row, terse reasoning per item keeps quality high and the per-item cost low across millions of calls.

Agentic Workflows

Keep each reasoning turn short inside multi-step agents so the running context stays compact and every tool decision is cheap, which matters most over long task horizons.

On-Device and Edge

Where compute and memory are tight, shorter reasoning traces reduce the work per query, making step-by-step reasoning more practical on constrained hardware.

Streaming UX

A short trace finishes streaming quickly, so the user sees the reasoning and the answer almost together instead of watching a long paragraph scroll by first.

Where Chain of Draft Fits

Part of the efficient-reasoning shift that follows Chain-of-Thought

Chain-of-Thought Reason Out Loud Explicit verbose steps
Chain of Draft Reason in Shorthand Terse minimal drafts
Sketch-of-Thought Adaptive Sketches Route among terse styles
Token-Budget Reasoning Budget the Length Cap tokens per problem
Chain These

Reach for Chain of Draft when you already trust Chain-of-Thought for a task but the verbose traces are too slow or too expensive at scale. Keep the same problems and the same step-by-step discipline, then simply add the minimal-draft constraint to recover most of the cost.

Reason for Less

Try adding a minimal-draft constraint in the Prompt Builder, or explore how Chain of Draft connects to the wider family of efficient-reasoning techniques.