Audit whether a support response actually answers the customer
Compare a draft response with every customer question, fact, constraint, and requested outcome
7 min setupTested with:ChatGPTReviewed: 2026-08-28
1
Add your context
Your text stays in this browser. AILesson Prompts does not send it to a model or server.
2
Your prompt
Unfilled fields remain visible as placeholders, so you can still copy and edit the prompt
Audit whether the draft support response truly answers the customer using the supplied facts and policy.
Latest customer message, relevant history, attempts, answers, requested outcome, and channel:
[customer]
Exact response including subject, links, attachments, promises, questions, and signature:
[draft]
Known status, approved steps, policy, authority, timing, escalation, privacy, tone, and disclosures:
[truth]
Decompose the customer message into explicit questions, implicit decisions required to proceed, requested outcomes, constraints, supplied facts, emotional or business impact, and safety signals. Build a coverage matrix mapping each item to the exact response text. Rate it answered, partially answered, unanswered, contradicted, redundant, or unsafe, with evidence. Then audit factual correctness, completeness, order of action, clarity, empathy tied to actual impact, effort imposed on the customer, repeated questions, authorization, promises, privacy, accessibility, escalation, and whether links or attachments really exist. Do not reward politeness that avoids the question. Distinguish blockers from optional style improvements. Recommend the smallest revision operation for each blocker; only rewrite if requested. End with a send, revise, hold, or escalate recommendation and human checks.
Private by defaultPrompt assembly happens locally in your browser. Avoid placing confidential information into any AI service unless your organization allows it.
From input to outcome
A worked example
See how concrete context turns this recipe into a usable result
Actual input
Customer request and history
Maya: 'Both PDF exports have blank charts. I need the charts for a board meeting tomorrow at 10:00. I already restarted Chrome, sent screenshots, and confirmed Chrome 140. Can you fix it today or give me another way to export?' Workspace ID already recorded. No report data sent.
Draft support response
Hi Maya, sorry for the inconvenience. Please restart your browser, confirm which browser version you use, and send a screenshot plus your workspace ID. Our team is looking into it and should have it fixed soon. Thanks for your patience.
Verified policy and facts
Known INC-882 affects some Chrome PDF exports; root cause unknown. Verified workarounds: export PPTX or use Firefox for PDF. Next public update Aug 28 08:00 Asia/Shanghai. No permanent-fix ETA. Do not request password/report data. Existing screenshot, browser 140, restart, and workspace ID should not be requested again. Escalate if both workarounds fail.
Example output
Recommendation: hold and revise.
Coverage matrix
• “Can you fix it today?” — contradicted/unsafe. The draft says “should have it fixed soon,” but no permanent-fix ETA exists. Replace with the known incident status, unknown root cause, next public update time, and an explicit statement that a permanent-fix time cannot yet be promised.
• “Give me another way to export” — unanswered. This is the customer's immediate decision need. Add the two verified options in action order: export PPTX, or use Firefox for PDF; add escalation if both fail.
• Board meeting impact — partially acknowledged only by generic “inconvenience.” Name the risk to tomorrow's board materials without exaggerating.
• Already supplied facts — redundant burden. The draft asks again for restart, browser version, screenshot, and workspace ID, all already present. Remove all four requests.
Correctness and authority blockers: “fixed soon” is unsupported; “our team is looking into it” is broadly consistent but less useful than the incident and update information. The draft includes no unsafe data request, but repeated evidence requests delay resolution.
Minimum revision operations: replace generic apology with impact-specific acknowledgment; identify INC-882 without claiming confirmed root cause; put the two workarounds first; state next update and ETA uncertainty; add escalation condition; remove duplicate questions. Human check the date/timezone relative to send time and verify the status link if one will be added.
Why this works
1
A question-to-answer matrix reveals friendly responses that omit the decision the customer needs.
2
Blocker separation focuses review on harm and resolution before style polish.
Check the result
Is every explicit question and requested outcome mapped to exact response text?
Are facts, authority, timing, links, and actions verified?
Are repeated effort and safety issues treated as blockers rather than tone issues?
ChatGPT