review-replies

---
name: review-replies
description: Turns unanswered reviews into publish-ready replies written for the next reader, and hands bulk drafting to Codex. Use when reviews sit unanswered.
---

# Answer every review in a way that sells the next customer

A voice guide, model replies per rating, and a reply drafted for every unanswered review.

## Claude does

1. Ask for: business name, what it sells, the town, who signs replies, and the review export (date, rating, reviewer first name, text). Ask what two or three things the owner wants future readers to know, in his words.
2. Refuse to guess. A review naming a person, booking, refund or date absent from the export is marked NEEDS OWNER and queried, never filled in.
3. Bucket it: 5 star, 4 star, 1 to 3 star, rating with no text. Report counts, average, and how many are unanswered.
4. Work oldest first. BrightLocal 2026, 1,002 consumers: 89% expect a response, 19% the same day, 81% within a week. Over seven days is overdue.
5. Write the voice guide: sentence length, banned words, and the short list of facts a reply may state (hours, service area, how to book). Nothing outside it is usable.
6. Write three model replies per bucket on one shape: thank by first name, name what they mentioned, add one true selling fact the next reader needs, invite the next step, sign with a real name. 35 to 60 words at 4 and 5 star, 50 to 90 at 1 to 3.
7. At 1 to 3 star: acknowledge, give a direct contact, never argue, never restate the complaint in the opening line, which search results show. Ban the superseo anti-slop vocabulary, and any two replies opening alike.

## Then Codex does

Codex applies the approved guide, model replies and banned list to every remaining row, one draft each, then scans hundreds of drafts for duplicate openings, banned words and word counts out of range. Bulk pattern work, no judgement.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Fill in first: {BUSINESS_NAME}, {SIGNER_NAME}, {REVIEWS_FILE}, {VOICE_GUIDE}, {MODEL_REPLIES}, {BANNED_LIST}.

## Claude checks

1. Every draft names something the reviewer wrote. Reject any that would fit a different review.
2. No draft states a fact outside the approved list.
3. No duplicate openings, no banned words, counts in range, every reply signed.
4. From the research: never quote ranking statistics at a customer, and never use the Whitespark category percentages, absent from its own page. That page scores ratings 181 and reviews with text 170.
5. Reject and rerun: over one failure in twenty sends the batch back, failing rows named. Never patch by hand.

## Rules

- Answer all of them. 80% say they are likely to use a business that replies to every review.
- Same day where possible, one week at the latest.
- Never ask for a rating, offer anything for a review, or write one.
- Never publish a customer's order number, address, phone or email.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude drafts the whole batch itself, twenty at a time, and says so.

## Built from

- BrightLocal Local Consumer Review Survey 2026, https://www.brightlocal.com/research/local-consumer-review-survey/ (1,002 consumers, February 2026).
- Whitespark 2026 Local Search Ranking Factors, https://whitespark.ca/local-search-ranking-factors/ (47 experts, 187 factors).
- inhouseseo/superseo-skills, https://github.com/inhouseseo/superseo-skills (305 stars, api.github.com).

# Answer every review in a way that sells the next customer - Codex task

You are drafting public replies to real customer reviews. Everything needed is in this file and the files it names. Do not search the web and do not add facts of your own.

## You are given

- {BUSINESS_NAME} - the business the replies come from.
- {SIGNER_NAME} - the real person whose name signs every reply.
- {REVIEWS_FILE} - reviews as CSV or text: date, star rating, reviewer first name, review text, and a status field marking rows already answered or NEEDS OWNER.
- {VOICE_GUIDE} - approved tone rules and the complete list of facts a reply may state.
- {MODEL_REPLIES} - three approved examples per bucket: 5 star, 4 star, 1 to 3 star, rating with no text.
- {BANNED_LIST} - words and phrases that must not appear.

## Produce

1. `replies.csv` - one row per unanswered review: `row_id, date, rating, reviewer, bucket, reply_text, word_count, status`. Status is READY or NEEDS OWNER.
2. `replies.md` - the same replies grouped by bucket, oldest first, each under a heading giving date, rating and reviewer, ready to copy one at a time.
3. `checks.txt` - four scans over replies.csv: duplicate opening sentences, banned words, word counts out of range, and replies mentioning no detail from their own review text.

Every reply follows one shape: thank by first name, name the specific thing the reviewer mentioned, add one true fact from {VOICE_GUIDE} the next reader needs, invite the next step, sign with {SIGNER_NAME}. Word counts: 35 to 60 for 4 and 5 star, 50 to 90 for 1 to 3 star, 25 to 45 for ratings with no text.

## Rules

- Only facts from {VOICE_GUIDE} may appear. If a reply cannot be written without inventing a detail, set NEEDS OWNER and leave reply_text empty.
- No two replies share an opening sentence, across the whole file, not just within a bucket.
- Nothing from {BANNED_LIST} appears anywhere.
- For 1 to 3 star: acknowledge, offer a direct contact to continue privately, never argue, never blame the customer, never restate the complaint in the opening line.
- Never publish an order number, address, phone number, email or any other personal detail of the reviewer.
- Never ask for a rating, offer anything in exchange for a review, or write a review.
- Skip rows already answered. Do not rewrite them.
- British English. No em dash characters.

## Return

Print, in order: rows read, replies written, every NEEDS OWNER row_id with its reason, the four scan results with counts, and the full text of the three longest and three shortest replies.