Post-booking retention engine

Is it for you?

Built for: operators with about 1,500 or more bookings a season on a booking platform with an API (FareHarbor, Peek, Checkfront, Cloudbeds or similar), plus a developer or vendor. A multi-day tour company or a 20-room lodge fits. You also need 2.a (the honest picture) and 2.b (approved facts) done first, because message A is built from them.

Smaller? The post-booking calendar from the session (five messages, one job each) gets most of the benefit with no code. The demo below still teaches the method.

Post-booking message rules

Write these down before the first draft. The reference code builds them in.

• Policy choice Nothing sends without a person. Drafts land in an approval queue as PENDING. The reference code has no sending function at all, and a test fails if anyone adds one.
• Policy choice No competitor price matching. Do not match rates and do not chase every cancellable booking. Pick the most exposed bookings by a rule you can read.
• Recommendation Confidence first, discount last. Every treatment booking gets message A. At most one follow-up: B (one add-on) or C (lock-in offer). C exists only if the owner writes its terms.
• Recommendation Approved facts only. Drafts use your fact sheet, your approved add-on list and your written terms. Anything missing stays as [FILL] for a person to supply. Asking the AI to check its own draft is not verification.
• Policy choice Contractual No guest personal data in the pilot. Only booking id, dates, rate type, channel, party size and add-ons. Names, emails, phones and payment data stay in the platform. Check your platform's API terms and the model provider's data terms.
• Contractual Promotions to OTA guests. B and C are promotional, so by default they go only to direct bookings. Check your OTA supplier agreement (for example Viator's) and any hotel-desk agreement before contacting their guests with offers.
• Legal mandate Commercial email and texts. Promotional email needs an honest subject line, your postal address, a working opt-out and, unless the guest agreed to receive it, a clear note that it is an ad. Opt-outs must be honored (CAN-SPAM Act, 15 U.S.C. 7701 to 7713; 16 CFR Part 316). Marketing texts sent with automated tools generally need prior express written consent (TCPA, 47 U.S.C. 227; 47 CFR 64.1200). Ask your platform how its messaging handles this. This is not legal advice.
• Legal mandate Lodging prices. A lodge offer that shows a price must show the total price, including mandatory fees (16 CFR Part 464, in effect since May 12, 2025). The rule covers short-term lodging, not tours.
• Policy choice Safety wording has an owner. Weather plan B, sea state, glacier and wildlife lines are drafted by AI and owned by your captain or operations lead. No promised sightings.
• Recommendation Holdout guests lose nothing. They still get every normal confirmation, safety and weather message. Only the pilot's extra messages are held back.
• Recommendation Stop rule. Pause the pilot if complaints or opt-outs rise, and decide before you start who makes that call.

How it works

1. Pull. Each night, export open bookings. The adapter keeps an allow-list of columns, drops everything else and logs the dropped column names, never their values.
2. Score. A rule set you can read and change. Higher score means more likely to change or cancel inside the free-cancellation window. The top 10% are eligible.
3. Assign. Eligible bookings go at random to treatment or holdout, 50/50 for a pilot. The draw is seeded, stored once and never redone.
4. Draft. For treatment bookings only: message A, then at most one follow-up. Templates only (dry run), or a model API improves the wording. Local checks flag new numbers, banned phrases, promises and personal data.
5. Approve and send. A person fills each [FILL], edits, approves, and sends from the booking platform.
6. Report. Monthly in season: in-window changes and cancellations, treatment versus holdout, flexible and non-refundable separately, with 95% intervals and a "directional" label when the sample is small.

The starting score

A starting guess to test, not a prediction model. The weights live in rules.json; change them, note who changed what and when Policy choice

Three messages, and nothing else

DRAFT - PENDING OWNER REVIEW
Subject: Your 4-hour wildlife and glacier cruise on Tuesday, July 14, 2026: what to expect

Hello,

Thank you for booking the 4-hour wildlife and glacier cruise on Tuesday, July 14, 2026 at 9:00. Here is what to expect, so there are no surprises.

What to expect
- Heated enclosed cabin and two restrooms on board.
- Hot drinks are included. Lunch is not included.
- Meeting point: the small-boat harbor, a 12-minute walk from the cruise ship dock.
- Weather: The cruise runs in rain. It cancels only for a marine warning, with a full refund.

If your plans change
- Free cancellation until 48 hours before departure. Inside 48 hours, no refund.
- To review or change your booking, reply to this message or call [FILL: contact person] at [FILL: contact phone]. A person on our team will help you.

Our track record
- [FILL: track record]

See you at the harbor,
The Bay Light Day Cruises team

DRAFT - PENDING OWNER REVIEW
Subject: One idea for your day in Seward on Tuesday, July 14, 2026

Hello,

You are booked on the 4-hour wildlife and glacier cruise on Tuesday, July 14, 2026 at 9:00. Guests often ask what else fits the same day, so here is one option we know well.

Mile 12 Kennel summer sled-dog cart ride (partner)
- 90 minutes, 15 minutes by car from the harbor, ages 2 and up, runs in rain.
- Prices and times come from Mile 12 Kennel directly.
- Book with Mile 12 Kennel and mention referral code BAYLIGHT-K.

It is optional, and your cruise booking stays the same either way. Please check the age limits for your group.

Questions? Reply to this message or call [FILL: contact person] at [FILL: contact phone].

The Bay Light Day Cruises team
[FILL: postal address and opt out]

DRAFT - PENDING OWNER REVIEW
Subject: An option for your Tuesday, July 14, 2026 cruise

Hello,

You are booked on a flexible rate for the 4-hour wildlife and glacier cruise on Tuesday, July 14, 2026 at 9:00. Nothing changes unless you choose to.

If you would like to lock in your booking now, here is our offer:
FICTIONAL EXAMPLE TERMS: switch to our non-refundable rate at least 30 days before your trip and we add a printed wildlife guide for each guest. (The owner writes the real terms.)

If you do nothing, your current booking stays as it is. Free cancellation until 48 hours before departure. Inside 48 hours, no refund.

Questions? Reply to this message or call [FILL: contact person] at [FILL: contact phone]. A person on our team will help you decide.

The Bay Light Day Cruises team
[FILL: postal address and opt out]

Each [FILL] is a fact the fictional owner has not supplied yet. The pilot never guesses it.

What the demo shows: in the fictional data, the most exposed bookings came through an online travel agency. So the pilot drafted only message A for them, and held B and C back for contractual reasons. For many operators, the confidence note is the whole lever.

Build brief

Give this to a developer, a vendor, or a coding agent (Claude Code, OpenAI Codex or Gemini CLI). It is the Prompt Kit brief with three updates: a 50/50 holdout for the pilot, Wilson intervals in the report, and checks on every draft.

Design and build a post-booking retention pilot on [BOOKING PLATFORM].

Nightly: pull open bookings through the platform API (booking id, lead time, rate type, channel, party size, add-ons, price-alert flag if available). No guest names, emails or payment data leave the platform.

Score exposure with a transparent rule set I can read and change, for example: +2 lead time over 60 days, +2 flexible rate, +1 OTA channel, +1 no add-ons. For the top 10 percent by score, call a model API to draft one of three approved messages and nothing else: (A) a confidence note stating what to expect and the weather plan B, (B) one relevant add-on from the approved list, (C) a lock-in offer to convert to non-refundable on terms I supply. The model must use only the approved facts file.

Write drafts to an approval queue (a Google Sheet or email drafts). Nothing sends without a person. Randomly assign 20 percent of eligible bookings to a holdout that receives nothing, and store the assignment. Report monthly: in-window changes and cancellations, treatment versus holdout, as counts and rates, with flexible and non-refundable bookings shown separately. Do not match competitor prices and do not chase every cancellable booking.

Read a result

At the end of a month, type the counts from your report. The page shows each rate with a Wilson 95% interval and says how far to trust it.

The demo result comes from the fictional, simulated season in the download. It looks like a big gap, but with 14 and 13 bookings it is most likely noise. Use the sample size above for "confirmatory": below it, the read is directional.

Download the reference pilot

A small Python program that runs the whole method on 300 fictional bookings, with no keys and no internet. It scores, assigns, drafts A, B and C into an approval queue, reports with Wilson intervals, and runs the sample size. It has no sending code.

Unzip it, open a terminal in the folder, and type:

On a Mac, type python3 instead of python. The README walks a non-developer through every step, including your own data and optional model drafts.

Look inside first

• README.md Start here. Plain-language steps, rules and limits.
• sample_output/example-drafts.md One draft of each type, plus a bad draft and the flags the checks raise.
• sample_output/output/approval_queue.csv The approval queue from the demo, every row PENDING.
• sample_output/output/report.md The demo report: treatment versus holdout with intervals.
• rules.json Scoring points, holdout share and seed, message order, banned phrases.
• approved_facts.md The only facts a draft may use (fictional).
• approved_addons.csv Partner add-ons with "fits when" rules (fictional).
• templates/A · B · C A, B and C wording with blanks.
• data/baylight-bookings-mock.csv 300 fictional bookings, no personal data.
• pilot.py The program.
• Prompt-Kit-3x3.html, section 2.c The original brief in the printable kit.

Platform connections (FareHarbor, Peek, Checkfront, Cloudbeds) are empty stubs: their APIs are not verified here. Optional model drafting works with Claude; the OpenAI and Gemini options are not tested against a live API. Effort and cost are estimates.

Before this rung

2.c is the top rung of Pain 2. Message A is only as good as the facts behind it: 2.a The honest picture gives you the "what to expect" lines, and 2.b Grounded answer assistant gives you approved facts that cite their source.