AI Guest Message Drafts: How Watchtower Reads Your Guidebook to Reply for You

If you host a vacation rental, you've answered the same questions 50 times. "What's the wifi?" "Can I check in early?" "Is parking included?" "Where do I leave the keys?" The answers don't change between guests — but typing them all out, every time, takes hours per month that you don't get back.

There are plenty of "AI for Airbnb" tools on the market, and they fall into two categories. Some auto-send replies on your behalf the second a message lands — fast, but guests increasingly notice the bot tone, and a wrong answer to a refund question is a five-star review you'll never get. The others require you to hand-craft a system prompt for every property, copy-pasting your wifi password and check-in instructions into a giant text blob the AI reads on every request. That's a maintenance nightmare the moment you change your wifi or your house rules.

A Different Approach: Knowledge Base From Your Guidebook

HostMoat's Watchtower is built on a simpler premise: you already wrote the answers when you set up your digital guidebook. Your wifi network and password live in the guidebook. Your check-in time, checkout time, and parking instructions live in the guidebook. Your house rules, your contact list, your local recommendations — all of it. That structured content is the AI's knowledge base. There's no separate prompt to maintain. Update your guidebook, and the AI's answers update automatically on the next request.

When a guest asks "what time is check-in?", Watchtower reads the check-in/check-out section of your guidebook and answers with your actual times. When they ask about pet policies, it reads your house rules section. If the answer isn't in any guidebook section, the AI is instructed to say it'll check with you and follow up — not to guess. That's the difference between a useful drafting tool and a hallucination machine.

Always Reviewed, Never Auto-Sent

Here is the design rule that everything else flows from: Watchtower never sends a reply on your behalf. There is no auto-send mode. Every draft lands in a review panel above your reply composer with three buttons — Send as-is, Edit and send, or Discard. Most drafts ship within seconds of you reading them. The point isn't to remove you from the conversation; it's to remove the typing.

This matters for two reasons. First, the model occasionally gets something subtly wrong — a date, a rate, a policy edge case — and you'll catch it during the half-second skim before sending. Second, your voice matters. Guests can tell the difference between a thoughtful host reply and a bot blast. Watchtower drafts in your tone (configured per property) and in your voice — but you still hit Send.

Bring Your Own Key — Free

On the free tier, you connect your own OpenAI API key, your own Anthropic / Claude key, or point Watchtower at a self-hosted Ollama server via a Cloudflare Tunnel. You pay the AI provider directly — HostMoat takes nothing on top, and your usage stays between you and them. Most owners spend $1-5 per month on OpenAI for typical guest-message volume.

Pro+ subscribers get HostMoat's built-in AI included with their subscription, no API keys required. The built-in option runs on Gemma 4, hosted by HostMoat, and there's no per-message cost beyond the flat Pro+ rate.

In all cases, your API keys are encrypted in Supabase Vault on the server side. The Settings page only ever shows a "Configured ✓" status — the browser never sees the actual key value after you save it. To rotate, click Replace and paste a new one; the old one is overwritten in Vault.

Customize the Voice and Tone

In Settings → Watchtower AI, set a Tone field with whatever voice fits your brand: "warm and casual", "professional and concise", "friendly with a Southern drawl" — anything natural-language. The AI will follow it. Add Owner Notes for context the guidebook doesn't cover ("if guests ask about late checkout, offer it for $30 when we don't have a back-to-back booking"). The AI uses these as soft instructions on top of the structured guidebook facts.

For full control, the Advanced Custom Prompt textarea lets you override the entire prompt — useful if you want a very specific voice or response structure. Use the {{message}} placeholder to mark where the guest's message should appear; if you forget, Watchtower appends it at the end with a delimiter as a safety net.

Test Before You Ship

The Settings page has two panels that demystify what the AI is actually doing. "What Watchtower knows" shows you the exact knowledge base text the model sees, broken down by source: which property fields it has, which booking-site content, which guidebook sections. Each item has an "Edit there" jump link to the source page so you can fix gaps as you find them. "Try a sample question" lets you type any hypothetical guest message and see the AI's response without sending anything — useful for confirming the tone is right or comparing providers before going live.

When Not To Use AI Drafts

Watchtower is a drafting tool, not a substitute for your judgment. If a guest is upset, asking for a refund, or in the middle of a complicated multi-message exchange, type the reply yourself — the situational nuance matters more than speed. The same goes for first-time guests asking detailed questions about availability or pricing edge cases; these are conversations where your voice and flexibility convert bookings, and a draft can flatten the relationship.

The right use case is the dozens of routine messages: confirmations, instructions, simple questions, friendly checkpoints. Those are the ones eating your evenings, and those are the ones Watchtower is built for.

The Bottom Line

AI in guest communication only works if it's grounded in your actual property data and if you stay in the loop. Watchtower is opinionated about both. Your guidebook is the knowledge base, your tone shapes the voice, and you approve every draft before it ships. No prompt engineering, no auto-send, no surprises.

Free tier: bring your own OpenAI or Claude key, pay them directly. Pro+: built-in AI included. Either way, every draft routes through your inbox for a final pass — because the host's judgment is the feature, not the model.