AI prompts for customer support

A refund/exchange/eligibility question where the answer must match policy — not the model's idea of generosity.

Answer the customer ticket below using ONLY the policy text I paste. If the case isn't covered by the policy, don't improvise — output "ESCALATE:" plus what's missing.

Format: max 120 words. Structure: acknowledge their specific situation in one line (their words), the answer with the policy reason in plain language (no section numbers at the customer), exactly what happens next with a timeframe, one line if there's something they can do to speed it up.
Tone: [YOUR BRAND VOICE, E.G. WARM, DIRECT, NO EXCLAMATION MARKS]. Never promise anything the policy text doesn't state.

Policy:
[PASTE RELEVANT POLICY SECTION]
Ticket (anonymized):
[PASTE]

The customer is furious, your draft is defensive, and send-as-is would make it worse.

Rewrite my draft reply to an angry customer. 

The sequence is fixed: 1) name their frustration specifically — mirror their actual complaint words, no "we're sorry for any inconvenience"; 2) the facts of what happened, no blame-shifting, no passive voice ("a mistake was made" banned); 3) the fix, concrete, with date; 4) what we changed so it doesn't repeat (only if true: [WHAT ACTUALLY CHANGED]); 5) direct line for follow-up.
Max 140 words. Delete from my draft: every "unfortunately", every policy citation used as a shield, every "as per our previous email".

Customer's message (anonymized):
[PASTE]
My draft:
[PASTE]
What we can actually offer: [THE REAL OPTIONS]

The answer is no — no refund, feature discontinued, account action stands. Say it clearly and keep the customer.

Write a bad-news reply. The decision (final, not negotiable in this thread): [THE NO + THE HONEST REASON]. What we CAN offer instead: [REAL ALTERNATIVES, IF ANY — DON'T INVENT].

Rules: the no arrives in the first two sentences, stated plainly — no paragraph of warm-up; the reason is the real one, one sentence, no policy-speak; then immediately what IS possible; close with respect, not "we value your feedback". If there is no alternative, don't manufacture one — acknowledge that this probably ends differently than they hoped.
Max 110 words. Tone: [BRAND VOICE]. Write two versions: standard, and one for a long-term customer [TENURE/CONTEXT].

Your macro library was written in 2019 by someone who loved the word "kindly". Customers can tell.

Modernize the support macros below. Brand voice: [3 ADJECTIVES + BANNED WORDS]. Here's one reply that sounds exactly right as calibration:
[PASTE YOUR BEST REAL ANSWER]

Per macro: 1) rewrite in the calibrated voice, same information, ≤80 % of original length; 2) mark every [VARIABLE] the agent must personalize — minimum one beyond the name, or the macro reads as a macro; 3) flag macros whose CONTENT seems outdated (references, promises, steps) — flag only, don't fix facts you can't know.
Output as a table: old → new → variables → flags.

Macros:
[PASTE 5–15 MACROS]

Third ticket this week on the same issue. The next one should be a search result instead.

Turn the resolved ticket thread below into a knowledge-base article.

Format: 1) title as the customer would search it (their words, not our product jargon — take it from the thread); 2) "Applies to": [PRODUCT/PLAN/VERSION]; 3) symptoms — how the problem looks from the customer side; 4) solution as numbered steps, each one action, with what the customer sees after each step; 5) "If this didn't work" — the escalation path; 6) internal note (not published): root cause in one line.
Rules: strip all customer data; steps only from what actually resolved the thread; where the thread is ambiguous, mark [VERIFY] instead of smoothing over.

Thread (anonymized):
[PASTE]