Rephrase Before Responding

Practice

I’m outsourcing my thinking and getting dumber

Problem

You ask for something; the AI delivers something adjacent — technically responsive to your words, aimed at a different intent. You spot it three paragraphs in, or worse, after the work is merged. The mismatch was present in the first seconds of the exchange; nothing surfaced it until the cost had compounded. And each time you let an adjacent answer slide because it’s “close enough,” your own definition of what you actually wanted gets a little blurrier.

Mechanism

Before executing, the AI restates the request in its own words — expanded, disambiguated, with its reading of the intent made explicit — and the human confirms or corrects. Misalignment gets caught at the restatement, when it costs one sentence, instead of at delivery, when it costs the work.

Two distinct effects:

Machine-side (documented): models answer their own rephrasing of a question better than the raw question — rephrasing surfaces ambiguities the model would otherwise resolve arbitrarily (the “Rephrase and Respond” result, cited above).
Human-side (the reason this sits in this category): the confirmation step forces you to check the restatement against your intent — which means you must actually hold your intent precisely enough to check something against it. That small repeated act is thinking retained. Skipping it — letting the AI’s interpretation stand unexamined — is thinking outsourced, one exchange at a time.

How to apply (you)

For any request where misunderstanding is plausible and rework is expensive, ask for the restatement before execution: “before you start, say back what you understand me to want.”
Read the restatement against your intent, not for plausibility — the question is “is this what I meant?”, not “does this sound reasonable?” Those diverge exactly when it matters.
Correct precisely: name the difference between what it said and what you meant. Precise corrections teach the AI your vocabulary for next time.
Don’t ritualize it on trivial requests — the restatement gate is for requests with divergence room.

What your agent does with this

For non-trivial or ambiguity-bearing requests:

Before executing, restate the request expanded: what is being asked, for what purpose, with what constraints — in your words, not an echo of the human’s.
Where the request is ambiguous, don’t silently pick a reading: state the reading you chose and mark it as chosen (“I’m reading X as meaning Y”).
Proceed on confirmation; on correction, update and restate the delta in one line.
For trivial unambiguous requests, skip the gate — restating “fix the typo” is ceremony.

How it sounds

The narration your agent uses so you can see the technique running — the same words you just learned here:

“My reading of the request: {restatement}. Flagging one interpretation I chose: {ambiguity} → {chosen reading}. Correct me or I’ll proceed.”

How you know it’s working

Human: “that’s not what I meant” arrives at the restatement, not at delivery. If you’re still discovering misalignment in finished work, the gate is being skipped or the restatements are echoes.
Agent: your restatement contains at least one thing the human’s literal words didn’t say — a purpose, a constraint, a chosen reading. A restatement that only reorders the human’s words verifies nothing.

Failure modes

The echo — restating by shuffling the request’s own words. It passes every check while verifying nothing; the restatement must expand, or it’s noise.
Rubber-stamp confirmation — the human skims “yep, go.” The gate’s value is the human’s actual check; confirming without reading converts the practice into latency.
Restating everything — applying the gate to trivial requests trains both sides to stop reading the restatements. Reserve it for divergence room.
Interrogation instead of restatement — replacing the restatement with a battery of clarifying questions the AI could have answered itself. State your reading and mark the chosen ambiguities; ask only what you genuinely cannot resolve.

Evidence & field notes

Evidence: real production use — the field notes below are generalized accounts of real instances. Sources:

Deng, Y., Zhang, W., Chen, Z., & Gu, Q. (2023). Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves. arXiv:2311.04205.

Machine-side, Deng et al. (cited) showed the mechanism experimentally: letting a model rephrase and expand a question before answering improves accuracy across tasks, because ambiguities that would be resolved silently and arbitrarily get resolved explicitly instead. In production use with frontier coding agents, the human-side effect dominated: restatement gates on multi-file changes caught wrong readings of scope (“refactor this” read as “redesign this”) before execution repeatedly. The practice’s own embedded form — a standing instruction that the agent confirm alignment by rephrasing understanding — has been observed to shift misalignment discovery from delivery time to request time across whole projects.

How to apply it in a prompt

The standing version, annotated:

“For any request where you could plausibly get my intent wrong: before executing, give me your reading of it in your own words — what I want, why I probably want it, and any interpretation you had to choose. Then wait.” — “in your own words” outlaws the echo; “why I probably want it” forces the restatement up to intent level, where the expensive misalignments live; “any interpretation you had to choose” surfaces the silent disambiguations, which are exactly the places adjacent-but-wrong answers come from. “Then wait” makes it a gate rather than a preamble the work steamrolls past.

“Skip this for trivial requests — if there’s no room for you to be wrong about what I mean, just do it.” — scopes the gate. Without the exemption, the practice degrades into ceremony on every message, and both sides stop reading — which is how a verification step dies.

Why this construction: the failure this prevents is invisible at request time and expensive at delivery time, so the prompt moves the check to where the failure is born. Every clause targets a specific degeneration of the practice (echoing, plausibility-reading, over-application) rather than asking generically for “confirmation” — which is how you get gates without checks.

Related techniques

Same family: Re-Read the Brief
Same family: Self-Ask Before Delegating