Rephrase and Respond: Rewrite First, Then Answer

Many times, the model answers off-topic not because it's dumb, but because your question itself is ambiguous, and it answers based on the wrong interpretation. Rephrase and Respond (RaR) is designed to plug this gap.

In one sentence: Let the model rephrase your question in its own words, fill in missing details, and then answer based on the rephrased version.

Why It Works

Between humans, if someone doesn't understand, they ask back: "You mean ... right?" Models don't ask back by default; they directly answer based on their own understanding—if that understanding is off, the entire answer is useless.

RaR forces the model to "confirm by paraphrasing" first. This paraphrasing often:

Exposes ambiguity: Clarifies vague parts.

Completes implicit conditions: Adds what you left unsaid.

Aligns understanding: Matches its interpretation with your intent.

Template

The simplest single-step version:

Please first rephrase my question below in a clearer and more complete way (fill in any implied conditions),
then answer the rephrased version.
My question: {your question}

The model will first output "Your question actually asks about ..." and then answer. You can immediately see if it understood correctly—if not, you can correct it right away, without waiting for a long irrelevant answer.

A Real Example

Original question:

How can I optimize this function?

Too vague—optimize what? Speed? Memory? Readability? Answering directly, the model can only guess one direction.

With RaR, the model first rephrases:

You want to know: Without changing the function's functionality, how to improve its execution performance (e.g., reduce time complexity, avoid redundant calculations)?

At this point, you realize "Oh, it thinks I want to optimize performance, but I actually want to optimize readability," and you can correct it in one sentence. That's the back-and-forth RaR saves.

Suitable Scenarios

Questions prone to ambiguity: Polysemy, unclear references, incomplete conditions.

High-cost tasks: Wrong answers are costly; better to confirm first.

Non-native language / imprecise wording: Let the model "translate" your question into a clear version.

Unsuitable Scenarios

When the question is already clear, it's a waste. "How to read a CSV file in Python" has no ambiguity; making it rephrase is just verbose.

It increases output length and cost. The extra rephrasing adds tokens. For batch or latency-sensitive scenarios, consider carefully.

Combining with Other Techniques

RaR addresses "input-side understanding alignment," while chain-of-thought addresses "output-side reasoning process." They handle different stages and can be stacked: first RaR to align understanding, then CoT to reason and answer.

It's also somewhat similar to Step-Back prompting—both involve "doing something else before answering"—but Step-Back abstracts principles, while RaR clarifies the question. Different purposes.

Summary

The essence of Rephrase and Respond is "let the model first confirm it understood correctly." The vaguer the question and the higher the cost of a wrong answer, the more worthwhile it is. If the question is clear, skip this step.