The swarm.
Six specialist agents plus a coordinator do the work that keeps FranchiseFrontline useful: watch every platform, route the bad reviews to the right human, draft brand-voice responses, follow up on leads, summarize the week. Every agent's prompt is published here so customers can read what we ask the machine to do.
Agents
review-aggregator
Pulls new reviews from Google Business Profile, Yelp, Tripadvisor, Facebook. Dedupes. Normalizes into one schema.
escalation-router
Triages each new review against the location's escalation rule. Decides who gets pinged first and when it climbs the chain.
response-drafter
Drafts a response in the brand voice. Platform-compliant. Never posts on its own.
brand-sentry
Watches brand-wide drift for HQ customers. Flags rating slides, guideline violations, platforms not connected.
lead-watcher
Classifies inbound marketing leads (operator, multi-unit, brand HQ, partner) and drafts first-touch replies.
social-poster (Phase 2)
Drafts and schedules outbound social posts across FB, IG, X, LinkedIn. Beta, opt-in.
queen-coordinator
The only agent that decides which other agent runs when. Generates no customer-facing output of its own.
Workflows
How the agents chain.
- daily-review-sync - every 10 minutes per active location, runs review-aggregator. Daily for trial accounts.
- escalation-cycle - on each new negative review, escalation-router fires, response-drafter generates, alert email or SMS goes out.
- weekly-recap - Mondays at 7am MT, brand-sentry + a recap composer generate per-operator and per-brand digests.
- lead-followup - on each marketing lead submission, lead-watcher classifies and drafts a reply for human review.
Hooks
Hard rules every agent obeys. Not optional, not negotiable.
- gate-outbound - no agent posts to a customer-facing surface without explicit human approval.
- scrub-customer-pii - reviewer names, emails, phone numbers must be scrubbed from prompts, logs, and any data we send to third parties.
- respect-platform-tos - never violate Google, Yelp, Tripadvisor, Facebook, or X terms of service. No fake reviews, no review-buying, no bulk auto-responses.
Memory
Shared knowledge the agents read from.
- brand-taxonomy - the categories of customer complaints (cleanliness, wait time, staff, hours, product, price, access).
- escalation-rules - per-location escalation tree shape and SLAs.
- response-style - brand voice settings, banned phrases, signature patterns.
Tools
External APIs the swarm reaches (no secrets here; auth lives in Cloudflare environment variables).
Why this is published.
You can read every prompt we feed the machine. If we ever start asking a model to do something we would not want a customer to see, this page would show it. That is the kind of accountability reputation software should have, given what it touches.