Chain of Draft Prompting: 4 Recipes That Ship (2026)

> Quick Answer (2026): Chain of Draft (CoD) keeps chain-of-thought's step-by-step reasoning but caps each step to a few words, so you stop paying for verbosity while holding accuracy. In the original 2025 paper it ran on as little as 7.6% of the tokens Chain-of-Thought used and still matched or beat it. Use it when reasoning helps but the word count is costing you money or latency. The one-line recipe: "Think step by step, but keep each step to at most five words. Put the final answer on the last line." Skip it for long proofs where dropping one step breaks the answer.

Chain of thought talks too much. You pay for every word of it.

Chain of Draft keeps the thinking and drops the padding. Same steps. A few words each. The claim, from the Chain of Draft paper (Xu, Xie, Zhao, He; 2025): CoD ran on as little as 7.6% of CoT's tokens and still matched or surpassed it on accuracy.

Tested on Anthropic Claude and OpenAI GPT models, July 2026.

Recipe 1: The one-liner

This is the whole technique.

Think step by step, but keep each reasoning step to at most five words.
Put the final answer on its own last line, starting with "Answer:".

Why it works: the word cap forces the model to emit only the load-bearing tokens. No restating the question. No warming up. Just the arithmetic or the logic, compressed.

Receipt: a word problem that ran roughly 200 tokens of reasoning under CoT collapses to about 30 under CoD. Same answer line. You paid for a sixth of the output.

Failure mode: too tight a cap on a genuinely long chain drops a step. Fix in Recipe 3.

Recipe 2: Make it parseable

If a script reads the output, pin the answer to a line you can grep.

Solve this step by step. Each step: at most five words.
After the steps, output one line that starts with "Answer:" followed by the answer.
Output nothing after that line.

Now you split on "Answer:" and take the tail. The draft steps stay in the response for auditing but never reach your parser.

Recipe 3: The gate (when NOT to draft)

CoD wins on arithmetic, classification, extraction, routing, short logic. It loses when every step is load-bearing: formal proofs, multi-hop legal or medical reasoning, code you will run. Compress those and the model drops a step, then answers wrong with full confidence.

Route it:

If the task needs more than about six reasoning steps, or every step is load-bearing, reason in full sentences.
Otherwise keep each step to five words. Then give the answer.

Recipe 4: CoD at scale (the cost move)

Batch jobs are where this pays. Classify 50k tickets, extract fields from 10k docs, and the reasoning tokens dominate the bill. The paper puts CoD at roughly 7.6% of CoT's tokens, so at any fixed output price that is close to a 90% cut on the reasoning you emit.

Keep an escape hatch so the hard tail does not go wrong quietly:

Draft each step in at most five words.
If you are not confident, write "LOW CONFIDENCE" on its own line, then reason in full sentences instead.

Now you pay CoD prices on the easy majority and full CoT prices only on the cases that need it.

Chain of thought vs chain of draft

Same skeleton. Different budget.

• Chain of thought: verbose, every step spelled out. Best when steps are load-bearing or you need an auditable trace. See our chain-of-thought recipes for the full baseline.
• Chain of draft: the same steps, a few words each. Best when reasoning helps but verbosity is pure cost.

CoD just stops paying for adjectives. When the risk is a wrong fact rather than a long chain, do not compress; reach for chain-of-verification instead.

For the general knobs, both Anthropic's prompting docs and OpenAI's prompting guide still assume verbose reasoning by default. CoD is the override for when the bill matters.

Ship it on your cheapest model first. If accuracy holds, keep it.

By Roo Iyer

Cost to test: about $0.01 for one CoD call on a small model. The CoT baseline you are comparing against runs a few times that on output tokens, which is the entire point.