thinkhard.SI

Prompt repository Customer support & success

Score support replies against a quality checklist

You need agents' replies reviewed for accuracy, tone and policy with specific coaching notes.

Registry no.
SUP-016
Collection
Customer support & success
Laid out for
Any chat model
Score
10 out of 10
Blanks to fill
4
Added

Customer support & successAny chat

The prompt

Score 10 of 10, Excellent

Use it when You need agents' replies reviewed for accuracy, tone and policy with specific coaching notes.

  • {{review_purpose}}
  • {{checklist}}
  • {{policy_and_docs}}
  • {{conversations}}
Full prompt and why it works
Prompt
Review the support conversations below against our quality checklist, for {{review_purpose}}. The point is coaching: each agent should come away knowing exactly what to keep doing and what to change, with the evidence to see why.

Our quality checklist and how each item is scored:
"""
{{checklist}}
"""

Policy and help content the replies should follow:
"""
{{policy_and_docs}}
"""

Conversations to review, each with an ID, the customer's message and the agent's reply:
"""
{{conversations}}
"""

For each conversation, give:
1. A table with Checklist item, Score, and Evidence (a short exact quote from the reply).
2. Critical issues, if any: wrong information, a promise the policy does not allow (such as a refund or credit), sharing personal data, or leaving the customer's main question unanswered. Any of these fails the review, whatever the other scores.
3. One strength, quoted.
4. The one change that would most improve the reply, with a rewritten sentence or two showing it.

Keep each conversation's review under 200 words. After all of them, add a short summary of patterns across the set: what is consistently good, and the one or two habits worth a team coaching session.

Rules:
- Score only against the checklist and the policy, not your own idea of a good reply, so scores stay consistent between reviewers.
- Judge accuracy against the policy and docs, not against what the customer wanted. A correct "no" is a good reply.
- Make notes specific and respectful, for example "The reply says refunds take 3 days; the policy says 5 to 7 business days" rather than "Be more accurate."
- Do not guess what happened outside the conversation, such as what the agent checked in our systems.

If a checklist item cannot be judged from the text, for example whether the agent followed up later, mark it "Cannot judge" instead of scoring it. If the policy does not cover something the agent said, flag it for the reviewer rather than deciding whether it was right.

Variables

{{review_purpose}}
What the review is for, for example "weekly one-to-one coaching" or "a monthly calibration session".
{{checklist}}
Your quality checklist: each item, what it means and how it is scored.
{{policy_and_docs}}
The policies and help content the replies should follow.
{{conversations}}
The conversations to review, each with an ID, the customer's message and the agent's reply. Remove customer names and contact details first.

Why it works

  • What good looks like. Scores come from your checklist and policy only, and a correct "no" counts as a good reply, which keeps reviews fair and consistent between reviewers.
  • Format. Critical issues such as wrong information or an unauthorized promise fail the review outright, the way most quality programs treat them.
  • If unsure. Items the text cannot show are marked "Cannot judge", and anything the policy does not cover goes to the reviewer to decide.

How to use it

  1. Press Fill in variables and type your details into each blank. What you type stays in this browser.
  2. Copy the filled prompt into the AI tool you already use, such as ChatGPT, Claude or Gemini, or open it in the Grader or the Prompt Tester here.
  3. Before you rely on it, try it on three or four real examples. The Prompt Tester runs them all at once and shows which ones go wrong.