Prompt AI to Write a Polite No That Still Lands

The usual failure when you prompt AI to write a polite no is not rudeness. It is softness. You get three sentences of appreciation, a refusal wrapped in conditional verbs, an apology, and an offer to revisit later that you did not authorise. The recipient reads it as an opening position and replies with a counter-offer, and now you are declining twice.

That is a specification problem rather than a drafting problem. "Polite" is the only constraint most people give, and a model optimising for polite will hedge, because hedging reads as polite in isolation.

Decide the shape before you ask for words

A refusal carries two payloads: the decision and whatever you owe the other side. The second one varies, and it determines the shape of the message. Pick one of four.

Most bad AI-written refusals are a one-reason no that has drifted into door-left-open because the model added a closing line nobody asked for. Naming the shape in the prompt prevents exactly that.

The prompt to write a polite no with AI

Write a short reply declining the request below.

SHAPE: one-reason no
RELATIONSHIP: client of two years, good standing
MY REASON (use this, do not invent another): capacity until mid-January
DECISION STATUS: final, not negotiable

Constraints:
- Under 120 words.
- State the decision in the first two sentences. Do not bury it.
- Exactly one reason. Do not add a second.
- No apology beyond one short phrase, and never apologise twice.
- No "unfortunately", no "I'm afraid", no conditional verbs about the
  decision itself. "I can't take this on" not "I don't think I'd be able to".
- Do not offer an alternative, a referral or a future date unless I
  listed one above.
- Do not restate their request back to them at length.
- Match a plain, direct register. No warmth padding.

REQUEST:
---
[paste here]
---

Half of that prompt is prohibitions, which is worth a caveat. Anthropic's prompting best practices advises the opposite as a default, to "tell Claude what to do instead of what not to do", offering "your response should be composed of smoothly flowing prose paragraphs" in place of "do not use markdown".

That is good advice and it applies here. The positive form of most of the list above is one line: "State the decision in the first two sentences, give exactly one reason, close with a single courtesy, and stop." Lead with that. Keep the prohibitions only for the habits that have no positive equivalent, which in a refusal are the hedging verbs and the unauthorised offer, because "do not offer them a future date" cannot be phrased as something to do instead. Capping the word count kills the cushioning paragraph. Forbidding conditional verbs about the decision is what turns a negotiable no into a final one. And the rule against unauthorised alternatives stops the model volunteering your January.

Why "do not invent a reason" matters

A model with no stated reason will supply a plausible one, and a plausible invented reason is a specific hazard in a refusal. You may be stuck defending a constraint you do not have, or stating something your recipient can check and disprove. Supply the reason, or choose the bare-no shape and supply none.

The related trap is reason inflation. Three reasons read as weaker than one, because the reader hears someone assembling a case rather than reporting a decision. One reason, stated once, is a fact. Three reasons is an argument, and arguments have counter-arguments.

When the no has consequences

Some refusals need more than brevity.

Declining a client's idea rather than their request is a different job, because the useful content is the why. We cover that case separately in telling a client their AI project idea will not work.

Declining on price is a negotiation, not a refusal, and should be shaped as a counter-offer with a number in it. The mirror-image skill is in prompting AI to negotiate a lower price with a vendor.

Declining something you promised earlier needs an acknowledgement of the change before the decision, which is closer to the structure of an incident update than a standard no.

Length is the constraint that does the most work

If you add only one rule to a refusal prompt, make it the word limit. A refusal gets worse as it gets longer, for a reason that has nothing to do with AI: every extra sentence is a new reason, a new hedge or a new offer, and all three weaken the decision.

Under 120 words leaves room for the decision, one reason and a civil close. Past roughly 200 words the message starts to read as a justification, and a justification invites a rebuttal. This is also why models produce such poor refusals by default. Nothing in a request for politeness implies brevity, so the fluent move is to keep reassuring the reader.

A useful test on the draft: delete every sentence that is not the decision, the reason or a courtesy, then see whether the message still works. It almost always does, at about half the length. Ask for that version directly next time.

Read it once for one thing

Before sending, check a single question: can the reader reply with a counter-offer that is still on-topic? If yes, and you did not want that, the decision is not stated firmly enough. Usually the culprit is one of three phrases, "at the moment", "right now", or "currently", any of which converts a no into a not-yet.

Then check that the voice survived. Models flatten toward a neutral business register, and a refusal that does not sound like you reads as a form letter, which is its own kind of insult in a long relationship. Supplying two sentences of your own writing as a register sample fixes most of it, and the broader approach to controlling output shape is in our guide to writing prompts that hold up.

Common questions

Why do AI-written refusals sound so apologetic?

Politeness is the only constraint in most prompts, and apology is the cheapest way to signal it. Cap the apology explicitly and the tone problem mostly disappears.

Should I give a reason when I decline?

Only when the relationship warrants one and the reason is true and not worth arguing about. A reason you cannot defend is worse than no reason.

How do I decline without closing the door for good?

Choose the door-left-open shape and name the condition precisely, such as a month or a specific change. Vague openness is what produces repeat asks.

Can I use the same prompt for a job rejection?

The shape carries over, the obligations do not. Hiring refusals often come with legal and record-keeping requirements, so check what your jurisdiction and your own policy require before sending anything.

Is it worth having AI write a two-line no at all?

Sometimes, for the awkward ones you have been avoiding for three days. The value is not speed, it is getting past the blank page on a message you do not want to write.