Cold Outreach Engine

• CF Worker
• D1
• Apollo / Sunbiz / UCC APIs
• Claude
• Resend
• Cal.com

Working production examples. Read the code. Steal the pattern.

• eFin Capital pipeline MCA broker stack — outreach + intake + UW · same pattern

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

1. 09:00–09:30 Pick a vertical YOU need to sell into.

   For Cafecito AI: Miami HVAC, dental, law firms. The principle: build the outreach engine for a vertical we've demonstrated we can serve. Eat your own dog food on Day 1, then sell the engine to others on Day 30.

2. 09:30–10:30 Pull the data: UCC + Sunbiz + Apollo for 50 prospects.

   Sunbiz active filings (free), UCC liens for filtering, Apollo for personnel + email enrichment. Manual cherry-pick to 50 high-quality. Quality > volume.

3. 10:30–11:30 Build the enrichment pipeline.

   For each prospect: scrape their site, pull last 5 Google reviews, check for recent funding, check for new hire signals. The richer the context, the more specific Claude can be.

4. 11:30–12:30 Build the first-touch generator.

   Claude reads enrichment + their typical service + a hand-picked observation → writes a 3-sentence cold email. Never templated. Each one references something only THEIR business has.

5. 12:30–13:30 Send 10 first-touches, manually approve each.

   Day 1 quality check: read every generated email before sending. Mark which are great, which need editing, which are template-tells. Update the prompt based on patterns.

6. 13:30–14:30 Build the 3-touch sequence.

   Day 0: cold pitch. Day 4: "still thinking?" with one new piece of value. Day 9: final no-pressure ask. After Day 9: archive. Three touches max. No "is this still a priority?" emails ever.

7. 14:30–15:30 Wire reply detection + Cal.com.

   Inbound reply → Worker classifies (interest / not now / unsubscribe / spam). Interest → reply with Cal.com link. Not now → schedule 90-day re-engagement. Unsubscribe → comply immediately.

8. 15:30–16:30 Build the dashboard.

   One screen: this week's sends, opens, clicks, replies, calls booked, by sequence step. Sortable. Export CSV. The dashboard is what you sell to the prospect once you've closed your own use case.

9. 16:30–17:00 Send the first 10 first-touches for real.

   After approving each, fire them. Set the sequence cron. Tomorrow you sell what worked.

• Data quality > volume. 50 hand-picked prospects beats 5,000 sprayed.

  Why: A 50-prospect list with deep enrichment generates 5-10 replies and 1-3 calls. A 5,000-prospect spray generates 30 replies, 0 calls, and a domain-reputation problem. Quality compounds; volume burns.

• Never use template language. "I noticed your website..." is a death sentence.

  Why: Every prospect has received that exact opening 50 times. Pattern-matches as automated outreach in 2 seconds. Pull from REAL signal: a press release, a new hire, a specific Google review.

• Three-touch sequence. Never seven. Never "is this still a priority?"

  Why: Sequences past 3 are universally hated and produce diminishing returns. The "is this still a priority?" email is the most universally complained-about sales pattern.

• Calendar link in the email signature only — never inline.

  Why: An inline "book a call here" button reads as a sales sequence. The link tucked in the signature is professional. Customers who want to book find it.

• Opt-out is one click, no form, immediate.

  Why: A list-unsubscribe header + one-click footer link is required by Gmail/Outlook in 2026. Beyond compliance, the brand cost of a multi-step opt-out is permanent.

• The first sentence cannot mention you, your company, or what you sell.

  Why: The first sentence either earns the second sentence or doesn't. Mentioning yourself or your product in sentence one is the universal disqualifier. Lead with an observation about THEM.

• Track domain reputation weekly. SPF/DKIM/DMARC all green or you're junk.

  Why: A single misconfigured DKIM record drops your inbox rate from 90% to 30% overnight. Set up Postmark/Mail-tester monitoring before sending; check weekly.

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

Build me a CF Worker enrichment pipeline for [VERTICAL] cold outreach.

Input: CSV of prospects (business_name, address, phone, owner_name, sunbiz_id).

For each row, enrich with:
1. **Site scrape**: hero text, service list, last blog post date, owner photo URL.
2. **Google reviews**: top 5 most recent + count + average rating + most common complaint.
3. **Recent signal**: scan for indicators — recent press release, "we're hiring," a LinkedIn post in last 90 days.
4. **UCC liens last 24 months**: filter list of recent secured debts.
5. **Personnel**: enrich via Apollo for the decision-maker — title, email pattern, LinkedIn.

Output JSON per prospect:
{
  "business_name": "...",
  "decision_maker": { "name": "...", "title": "...", "email": "...", "linkedin": "..." },
  "site_summary": "...",
  "services": [...],
  "reviews_summary": "...",
  "recent_signal": "...",
  "ucc_status": "...",
  "enrichment_score": "high|medium|low",
  "ready_for_outreach": true|false,
  "data_completeness": 0-100
}

Bindings: D1 PROSPECTS_DB, KV ENRICHMENT_CACHE (24h TTL on Apollo + Places), secrets APOLLO_KEY + GOOGLE_PLACES_KEY.

Output: complete worker.js, ready to run as one-shot CSV ingest or scheduled cron.

Write a 3-sentence cold email + 2-sentence WhatsApp follow-up to [DECISION MAKER NAME] at [BUSINESS NAME].

Enrichment:
- Decision maker title: [title]
- Industry: [industry]
- City: [city]
- Most recent specific signal: [e.g. "just hired a director of ops in March 2026"]
- Top services: [list]
- Reviews summary: [3 sentences]
- Owner persona inference: [warm / direct / busy / detail-oriented]

Our offer (one of):
- [Block 01: bilingual voice receptionist · $400/mo]
- [Block 02: smart intake form · $300/mo]
- [other Block]

Constraints:
- First sentence must NOT mention us or our product.
- First sentence must reference SPECIFIC signal from enrichment.
- Second sentence connects signal to a problem they likely have.
- Third sentence offers a tangible next step (Loom, sample, 5-min call).
- No emojis. No "circle back."
- 80 words max email; 30 words max WhatsApp.
- Sign with [SENDER FIRST NAME] from [SENDER COMPANY].

Output JSON:
{
  "email_subject": "...",
  "email_body": "...",
  "whatsapp_body": "...",
  "first_sentence_signal": "...",
  "confidence": "high|medium|low"
}

Generate Day 4 and Day 9 follow-up emails for cold outreach to [DECISION MAKER] at [BUSINESS NAME].

Original Day 0: [paste]
Reaction so far: [opened: yes/no, clicked: yes/no, replied: no]
Enrichment summary: [paste]

Day 4 (160 words max):
- Lead with NEW value. Options: a screenshot of competitor doing the thing; a 30-second Loom of our solution; a specific quote from one of their Google reviews + how the proposed solution would have prevented it; a one-paragraph case study.
- Connect to the original ask.
- Restate next step softly ("if useful, here's the 30-sec demo: [LOOM]").
- No "wanted to bump this up" language.

Day 9 (80 words max):
- Final touch. No-pressure energy.
- Open: "Last note from me on this — won't keep nudging."
- One-sentence value prop, in their language.
- One-sentence: "If you'd rather I check back in 90 days, just say '90.' If never, just say 'never.'"
- Sign + tiny calendar link.

Then archive after Day 9 unless they engage. Reply "90" → schedule re-engagement in D1. Reply "never" → opt-out permanently.

Output: subject line + body for each.

Build me two functions:

1. **Reply classifier**: takes inbound email reply and returns:
   - category: book | not_now | unsubscribe | hostile | not_dm | other
   - suggested_action: send_calendar | reschedule_90d | opt_out_immediate | escalate_to_human | archive_no_action
   - one-sentence response if appropriate
   - confidence: high|medium|low

   Examples:
   - "Sure, send me a time" → book
   - "Not interested right now" → not_now
   - "Stop emailing me" → unsubscribe
   - "I'm not the right person — talk to [NAME]" → not_dm + extract referral
   - "Maybe in Q3" → not_now + reschedule_90d

2. **Weekly domain reputation check** (cron Mondays 8am):
   - Use mail-tester.com API to score sending domain.
   - Check SPF / DKIM / DMARC valid.
   - Pull last week's metrics from Resend: sent, delivered, bounced, complained.
   - If complaint rate > 0.1% OR bounce rate > 5% OR domain score drops below 8/10: alert ops + auto-pause sends.
   - Otherwise: send small "all good" digest.

Output: both functions as ESM JS, plus sample alert email + sample weekly digest email.

Anchor on SDR economics. Junior SDR in Miami runs $60-80k all-in and produces ~150 personalized touches/month. The system produces 1,000+ at higher quality. Per-prospect cost: ~$0.30 (Apollo $0.10 + Claude $0.15 + Resend $0.001 + cron $0.001). At 1,000 prospects/mo marginal cost is $300 against $1k fee. Setup includes: vertical config, enrichment pipeline, prompt tuning per vertical, sequence design, dashboard, domain configuration. Does NOT include: dedicated outbound domain ($150 — client owns), CRM sync to HubSpot/Salesforce ($1.5k each), LinkedIn outreach integration ($2k). Eat-your-own-dogfood angle: ALWAYS lead with the version you used to find them.