AI Front Desk QA

You are building an AI Front Desk QA (Block 15 in the Cafecito AI new-hire playbook). Full reference at https://cafecito-ai.com/new-hire/blocks/15-ai-front-desk-qa. Read it. Use plan mode. Stop at every [GATE].

INPUTS YOU NEED FROM THE HUMAN (ask before doing anything else):
1. Prospect — phone-heavy business not yet ready for full voice automation (sells the pain before Block 01)
2. Decision-maker email
3. Source data: (a) anonymized call log export OR (b) public review pattern showing call complaints OR (c) synthetic data matching their likely failures

ENVIRONMENT (verify):
- Working dir: /home/eratner/cafecito-ai
- Cloudflare account: f7a9b24f679e1d3952921ee5e72e677e

SECRETS TO CONFIRM:
- ANTHROPIC_API_KEY (Claude — call classification)
- RESEND_API_KEY (weekly digest)

THE PLAN:

STEP 1 — RESEARCH + DATA (15 min, mostly waiting on human)
- Pull prospect site. Identify primary services + their stated hours + any after-hours messaging.
- Wait for the data input. If review-pattern: scrape last 30 reviews, filter for call-handling-related complaints. If synthetic: human provides 10-30 sample call records.
[GATE 1 — show data shape + review excerpt patterns, ask "proceed?"]

STEP 2 — CLASSIFICATION RUBRIC (10 min)
- Define the 5 failure categories: (1) missed revenue (caller wanted to book, no answer), (2) language mismatch (Spanish caller, English-only response), (3) unanswered question (caller had a basic question, no info given), (4) followup gap (call ended without scheduling/confirming next step), (5) other.
- For each: criteria + example + recovered-revenue estimate per occurrence.
[GATE 2 — show rubric + sample classifications on 3 records, ask "categories right?"]

STEP 3 — SCAFFOLD WORKER (10 min)
- Create /home/eratner/cafecito-ai/qa-<prospect-slug>/.
- Worker handles: POST /admin/ingest (uploads call records or scrapes reviews), GET /qa-audit/<slug> (hosted report page), GET /admin/digest (Sunday cron triggers Resend weekly summary).
- Bindings: D1 QA_DB.
[GATE 3 — show worker.js, ask "deploy?"]

STEP 4 — RUN THE CLASSIFICATION (15 min)
- Run Claude over each call/review record with the rubric. Output: category, recovered-revenue estimate, illustrative quote.
- Aggregate: total missed-revenue dollars, % of calls in each category, top 3 patterns to fix.
[GATE 4 — show aggregated results, ask "build the report page?"]

STEP 5 — REPORT PAGE (10 min)
- Hosted page at /qa-audit/<slug>: header with the headline number ("$X estimated monthly revenue lost to call-handling failures"), 5 category breakdowns with examples, top-3 recommendations (which Block from the catalog would fix each), CTA to book a 15-min walkthrough.
- Editorial style + their accent color if findable.
[GATE 5 — show URL, ask "ready for cold pitch?"]

STEP 6 — DRAFT THE COLD PITCH (5 min)
- Email subject: "We audited your phone handling. Here's what's leaking."
- Email body: 4 sentences, lead with the headline number, link the audit page, name $1k audit / $300/mo monitoring, one yes/no close ("want to walk through the top 3 fixes Tuesday?").
[GATE 6 — show draft, ask the human to send manually]

STEP 7 — WIRE WEEKLY DIGEST (5 min)
- CF Cron Trigger: 0 14 * * 0 (Sunday 9am ET) — runs the audit again on new data, sends Resend digest with the delta.
- The digest is the retention mechanism (similar to Block 05 monthly report pattern).

STEP 8 — SHIP THE SUMMARY
- Single-line: "[BUSINESS] QA audit at [URL] · headline $[N] missed/mo · pitch sent."
- Append to /home/eratner/cafecito-ai/qa-shipped.md.

DONE.

GUARDRAILS: NEVER claim recovered-revenue numbers without showing the math. Always label as "estimated based on industry averages." Cost ceiling: $3.