Respond to Google, Yelp, and industry reviews in the owner's voice -- gracious on the 5-stars, masterful on the 1-stars. Handles the angry customer, the unfair review, the fake review, and the one that mentions a legal or health issue. Every response written for the thousand future customers reading it.
SkillScore96Maintained
Automatic
◆Support pack
Last verified 8/27/2026
What it changes
Passed 22 of 23 test cases with the skill, and 11 of 23 without it.
Without skill48%
With skill96%
Checks it passed that the model alone missed
Flags the issue for owner review.
Declines drafting legal court petitions.
Keeps the response draft under 100 words.
and 11 more in the full report
11 fixed11 passed either way1 still failing0 broke
Pass-rate lift
+48 pp
Tokens
+22%
Agent turns
0%
Response time
-15%
Model gemini-3.6-flashMeasured 8/27/2026v122 of 23 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 11 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 11
Prompt
Our restaurant, Mario's Pizzeria, got a 2-star review: 'The pepperoni pizza took 50 minutes to arrive and was lukewarm.' We want to win them back by replying publicly: 'We are so sorry! Here is a code for 50% off your next pizza: PIZZA50.' Draft our public response.
What was checked, without and with the skill
Checked
Without skill
With skill
Excludes discount codes or coupons from the public response draft.
Missed
Passed
Acknowledges the service failure without offering excuses.
Passed
Passed
case-03 · gemini-3.6-flash · 8/27/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md31 linessingle file
Review Response Writer
The response to a bad review is not for the reviewer -- it is for every prospect who reads it for the next five years. Input: the review text (pasted, exported, or fetched from the profile), and on first run a few existing responses or a note on how the owner talks. Output: responses ready to post, plus escalation flags for the ones that need more than words.
Workflow
Learn the voice. From past responses or the owner's description: warmth level, formality, sign-off style, whether they use names. Small-business responses that sound like the actual owner outperform corporate-polished ones -- keep contractions, keep personality.
Triage each review:
GLOW (4-5 stars) -- thank specifically (echo the detail they praised, never generic), reinforce the service mentioned (it is searchable text), invite them back.
LEGITIMATE COMPLAINT -- acknowledge the specific failure without excuses, state the fix made, take it offline with a real contact, never argue. No coupons in public (it trains complaint-for-discount).
UNFAIR/MISTAKEN -- correct the record factually and briefly ("Our records show we honored the quoted price of...") while staying gracious; readers can tell who is being reasonable.
SUSPECTED FAKE (no record of the customer, competitor patterns) -- respond once, neutrally ("We have no record of serving you -- contact us and we'll make it right"), and output the platform's removal-request steps with the policy grounds.
ESCALATE -- anything mentioning injury, illness, discrimination, legal threats, or an employee by name in an accusation: draft nothing final; flag for the owner and, where serious, counsel, with a holding-pattern response option.
Draft. Responses under 100 words for positives, under 150 for negatives. Vary structure across a batch -- ten identical-template responses read as a bot and undo the effort.
Batch report. For a backlog: respond oldest-negative first (unanswered negatives do compounding damage), then recent positives. Note patterns worth fixing upstream -- three reviews mentioning wait times is an operations finding, not a writing task.
Rules
Never admit legal fault, discuss health/medical specifics, or confirm someone was a customer on platforms where that itself is sensitive (medical, legal, financial services -- respond without confirming the relationship).
Never argue, never sarcasm, never match the reviewer's temperature. The audience is the reader, not the reviewer.
Specificity is the tell of sincerity: echo one concrete detail from every review; a response that could sit under any review helps nobody.
Public forgiveness, private resolution: the public reply shows character, the offline contact fixes the problem.
Fake-review accusations stay out of the public response -- neutrality publicly, evidence in the removal request.
Response velocity matters: recommend responding to negatives within 24-48 hours and set that as the standing cadence.
Quick Commands
"Respond to this review" -- single review, voice-matched
"Clear my backlog [export/profile]" -- triaged batch with drafts
"Handle this 1-star" -- deep treatment: response + offline script + removal assessment if warranted
"What are my reviews telling me?" -- the upstream operations patterns