Review Response Bot

• CF Worker
• D1
• Google Places + Yelp APIs
• Claude
• Resend (daily digest)

A real prospect. A real demo. A real outbound message — all before 5pm.

1. 09:00–09:30 Pick a high-review-count business that ignores reviews.

   Wynwood restaurant, Brickell salon, Coral Gables dental office. Filter by 50+ Google reviews, owner-response rate < 30%. The lower their existing response rate, the bigger your before/after.

2. 09:30–10:30 Pull their last 50 reviews via Google Places API.

   Save: rating, text, author, date, owner-response (if any). This is your dataset for both voice training AND the demo.

3. 10:30–11:30 Train the voice on their About page + menu copy.

   Owner's About bio + sample social posts + any newsletter copy = voice corpus. Feed Claude these as system context. Test on 3 sample reviews — do they sound like the owner?

4. 11:30–12:30 Build the response generator.

   CF Worker endpoint that takes a review JSON + the voice corpus + tone (warm/direct/witty per owner) → returns a draft response. Match formality of the original review. Address by name only if the review used a name.

5. 12:30–13:30 Build the daily digest email.

   Resend cron at 8am: pull last 24h of reviews, generate drafts, build an email with each review + draft + one-click APPROVE / EDIT / SKIP. Approve fires the response back to Google Places API.

6. 13:30–14:30 Wire one-click approve.

   Each button is a signed URL → CF Worker endpoint → posts the response to Google. No login needed (token in URL). Token expires in 7 days. Log every action.

7. 14:30–15:30 Backfill demo: generate responses for their last 10 unanswered reviews.

   Don't POST them — generate the drafts. Build a one-page report: "Here are 10 of your reviews from the last 90 days you didn't respond to. Here's what I'd say in your voice for each."

8. 15:30–16:30 Pretty up the demo report as a one-pager.

   Hosted page at subdomain.cafecito-ai.com/demo/<owner-slug>. Their logo, their reviews, the AI-drafted responses side by side. This IS the cold pitch.

9. 16:30–17:00 Send the cold pitch.

   Email + Instagram DM. Subject: "10 of your customers wrote you in 2026 and you didn't answer." Link the demo report. Reply target: tomorrow morning.

• Never auto-post. Always one-click approve from the owner's email.

  Why: A response from a bot reads like a bot the moment it tries to address something specific. One bad auto-response wipes out the goodwill from twenty good ones. The 5 seconds of approval is the entire moat.

• Match the formality of the original review.

  Why: A casual "loved it!!!" gets a casual one-line response. A 200-word complaint gets a measured 4-sentence reply. Mismatched tone reads as scripted. Detect formality from word count + emoji presence + punctuation density.

• Address by name only if the review used a name.

  Why: Half of Google reviewers use a real name; half use "Sarah W." or just initials. Calling a "Sarah W." reviewer "Sarah" without confirmation feels presumptive. If the name is partial or absent, use "Hi there" or "Thanks for stopping by."

• Mention specific items / details the reviewer mentioned. Never generic praise.

  Why: "So glad you enjoyed your visit!" is the universal tell of a fake response. "Glad the bistec a caballo hit the spot — Marisol takes those pickles personally" is the response that gets quoted on Yelp as "owner who actually reads."

• Never offer a refund, comp, or remediation in writing without owner approval.

  Why: A bot offering "we'd love to make this right with a free meal" creates a binding offer the owner has to honor. Default the AI to "we'd love to make this right — please reach out to [owner contact]" and let the owner do the actual offer offline.

• Log every response with the original draft, edits, owner action, and timestamp.

  Why: For legal: "did the owner approve this response?" needs an answer. For improvement: which drafts got edited and how teaches the system the owner's preferences over time.

• Respond within 48 hours. The 48-hour window is the SEO + trust threshold.

  Why: Google's local-pack algorithm weighs response recency. Customer-facing trust drops sharply after 48h. Set the cron + the owner expectation at 48h max.

Drop these into Claude Code. Replace the [BRACKETED] fields with the prospect's details.

I'm building a review-response bot for [BUSINESS NAME] in Miami. I need to capture the owner's voice so the responses sound like them.

Below is the corpus of public writing from / about the owner — their About page, social media bio + last 20 posts, any newsletter intros, any quoted media interviews.

[paste corpus here — aim for 1,500-3,000 words total]

Analyze this corpus and produce:
1. **Voice profile** (under 200 words): tone register, signature phrases, words they use vs avoid, sentence length, emoji use, formality with strangers vs regulars.
2. **5 sample responses** to these dummy reviews:
   a. 5★ "Loved everything!!!"
   b. 5★ detailed review of a specific item
   c. 4★ with a small complaint about wait time
   d. 2★ with a billing dispute
   e. 1★ angry vague complaint, no specifics
3. **Style guide** (under 150 words): rules to give the response generator going forward.

Output as three labeled sections.

Generate a response to the following Google/Yelp review for [BUSINESS NAME].

Review:
- Rating: [1-5]
- Author: [name or partial]
- Date: [timestamp]
- Text: [full review text]

Voice profile and style guide:
[paste voice profile + style guide from previous prompt]

Constraints:
- Length: match review's length.
- Tone: match review's formality.
- Address: by name ONLY if the review used a real-looking name. Otherwise "Hi there" or platform-neutral opener.
- Specificity: reference at least one specific detail the reviewer mentioned.
- Never offer refunds, comps, or remediation in writing. Default to "please reach out to [OWNER_CONTACT]" for complaints.
- Never use: "circle back," "going forward," "team," "valued customer," "five-star experience," any emoji unless the original used emoji.

Output JSON:
{
  "draft_response": "...",
  "tone_used": "warm|direct|measured|apologetic",
  "addressed_by_name": true|false,
  "needs_remediation_offline": true|false,
  "confidence": "high|medium|low",
  "skip_recommended": true|false,
  "skip_reason": "..."
}

Skip if: incomprehensible, possibly spam, contains personal attack on a named employee, or mentions ongoing legal action.

Generate an HTML email layout for the daily review-response digest sent to [OWNER NAME] at [BUSINESS NAME] at 8am.

Constraints:
- Plain HTML, inline CSS (Gmail/Outlook safe), no JS, no external images.
- Mobile-first.
- Cream #FDF9F2, ink #1A1814, accent #B8472A.

Structure:
- Header: "Good morning, [first name]. [N] new reviews to look at."
- Summary: "[N] need a response · [N] you may want to skip · ~[N] minutes total."
- For EACH review (most recent first):
  - Reviewer name + star rating + date + platform.
  - Original review text (italic, indented).
  - Drafted response (regular weight, slightly bigger).
  - Three buttons: APPROVE (cream-on-accent, large), EDIT (text link), SKIP (text link). Each is a signed magic-link URL.
  - Confidence badge if low.
- Footer: "[N] reviews from the last 24 hours. To change frequency or pause: [link]. Sent by [BUSINESS NAME]'s review concierge."

Output: complete HTML email template with [PLACEHOLDERS].

Build me a one-page hosted demo report for [BUSINESS NAME]. Lives at demo.cafecito-ai.com/[business-slug].

Inputs:
- Their logo (or generate a clean text wordmark in their brand color)
- Their last 10 unanswered Google/Yelp reviews
- AI-generated responses in their voice
- The 48-hour-window stat
- Their current public response rate

Page structure:
- Above fold: "Hi [OWNER]. Here's what 10 of your customers wrote you in 2026 — and what we'd say back, in your voice."
- Stats banner.
- For each review: original (left), AI-drafted response (right), [APPROVE PUBLISH] teaser CTA.
- Footer: "Want this running every morning at 8am? $1,000 to set up, $250/mo. Reply to [EMAIL]." Cal.com embed.

Style: editorial, calm, scannable. Cream/ink palette. Fraunces headers, Inter body. Mobile-responsive.

Output: single-file HTML, no framework, inline CSS.

Anchor on SEO + repeat-visit value. Local-pack ranking weighs response rate; sustained 80%+ moves businesses up half a position on average within 90 days. Per-response Claude cost: ~$0.02. Google Places API: free up to quota. At 30-50 reviews/mo, marginal cost under $5/mo. Setup includes: voice training, digest setup, magic-link approval, demo report. Does NOT include: Yelp Fusion API ($300, requires partner approval), multi-location ($500 each), TripAdvisor/OpenTable adapters ($500 each). Upsell: combine with Block 07 (SMS reactivation) → $500/mo combined "Customer Loop."