The text bot told 13 people they were booked. The studio has 8 wheels.

"You're all set for Saturday at 2 PM! See you at the wheel."

That text went out 13 times from Wheelhouse Clay Co. (a made-up pottery studio). Tess owns it. Her Saturday open-wheel session has 8 wheels, at $45 a seat.

The bot was polite, fast and wrong. Nothing checked the seats before it said yes.

AI Prompts, Reviewed takes one prompt or workflow from How I AI, Claire Vo's show where guests share their screens and walk through how they really use AI. We run it on a small-business job, score it, and tell you what we'd change.

Credit: Based on "How to Build a Low-Code AI Concierge with Airtable, n8n, and Twilio," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by Andrew Mason (founder of Groupon and CEO of Descript at the time of recording) with Nabil Hyatt (partner at Spark Capital) in the episode "How I AI: Building a Real-World Business with Andrew Mason & Nabil Hyatt." Read the original workflow on How I AI. Watch the full episode. Guest site: none listed on the How I AI workflow page.

What the original does

Andrew built a text concierge for a board game club. A member texts something like "set up a game for me this weekend," and the system tries to form a group and reserve a table. It runs in 6 steps:

• Define one bounded job with clear rules for identity, consent, capacity, timeouts, cancellations and a human handoff.
• Keep records in Airtable: members, consent, availability, tables, requests and reservations, with explicit statuses.
• Route every text through n8n in fixed stages: receive, identify, check permission, plan, act, confirm and log.
• Give the agent small tools like "get member" and "reserve table," each checked on the server.
• Write a plain-English system prompt around those rules.
• Connect Twilio with opt-out handling and quiet hours, then test the ugly cases before any customer sees it.

The line we'd tape to the monitor: "Never invent records or claim a reservation before the tool succeeds." That is exactly the rule Tess's first bot broke.

The test

Everything here is made up. Tess wants members to text the studio to grab open-wheel seats, so she stops answering booking texts during class.

Run 1, the shortcut: Tess pasted one instruction into an AI texting add-on: "Answer texts from members and book them into open-wheel sessions."

Over one weekend, the bot confirmed 13 people for 8 wheels. It never looked at a seat count. It also replied to 3 members who had texted STOP months earlier, because nobody told it who had opted out.

On Saturday, 5 people arrived to find no wheel. Tess comped each of them a future session. At $45 a seat, that is 5 x $45 = $225 in comped seats, plus two annoyed reviews and a studio full of people standing around.

A bot that says "booked" without a booking is making promises in your name.

The rerun

Run 2: same studio, same members, our prompt below. Before any tech, the AI helped Tess write the rules on one page.

The rulebook said a seat exists only when the booking app shows it. The bot may read the seat count and request a hold, but it says "confirmed" only after the app replies with a booking number. When the session is full, it offers the waitlist or the next session. Anyone marked STOP gets no outreach at all.

Then the AI wrote 10 test texts, including a duplicate text, an unknown number, a STOP reply and a request for a full session. Tess ran all 10 on her own phone before turning the bot on.

That weekend, 13 members texted again. Eight got confirmed seats with booking numbers. The other 5 went on the waitlist and were offered Sunday. Four took it, so Sunday added 4 x $45 = $180. The 3 opted-out members got nothing.

Against the shortcut, that is $225 not given away plus $180 earned: a $405 swing on one weekend.

Verdict: 7/10

Strengths: The rules are excellent. One bounded job, records as the source of truth, small tools, consent checked before every send, and no confirmation until the booking really exists. The runbook of failure cases is the part worth stealing, even if you never touch n8n.

Weakness: It's built for someone comfortable wiring Airtable, n8n and Twilio with webhooks, retries and idempotency keys. Most shop owners already have a booking app and a texting tool. The original also assumes you will write the edge-case tests yourself, and that is the step owners skip.

Founder tweak: Write the one-page rulebook and the test texts first, let your booking app own the seats, and make the bot say "confirmed" only with a booking number.

Copy the MCAFax version

This is our own rewrite, not the original. Copy the original from Andrew Mason on How I AI.

I own [business type]. I want a text assistant that handles ONE job: [book open seats / fill a class / schedule pickups].

My booking or order system is [app name]. My texting tool is [tool].

PART 1. Define the job. Do not write the bot yet.

Ask me what customers text about, what the bot may do, and what it must never do.

Ask for capacity limits, cutoff times, cancellation rules and my quiet hours.

Repeat it back as a ONE-PAGE RULEBOOK and wait for my OK.

PART 2. Source of truth.

List every fact the bot needs (seats left, member status, opt-outs) and which system holds it.

The bot may only read those facts or request an action. It never stores its own version.

PART 3. Write the bot instructions.

Say "confirmed" only after the system returns a booking or order number.

When full, offer the waitlist or the next open slot. Never overbook.

Never text anyone marked STOP or opted out. Never share one customer's details with another.

If anything is unclear or a tool fails, stop and send it to me with a short summary.

PART 4. Test before launch.

Write 10 test texts: normal, duplicate, unknown number, STOP, full session, late cancel, typo, angry customer, two people booking the last seat, tool failure.

For each, show the reply I should expect. I will run all 10 before going live.

Paste into any AI chat, then run every test text yourself before a customer ever sees a reply.

Try this week

1. Pick the one booking or ordering task you answer by text most often.
2. Write the rulebook: capacity, cutoffs, cancellations, quiet hours and who has opted out.
3. Check that your booking app, not the bot, holds the seat count.
4. Run the 10 test texts and fix every reply that says "booked" without a booking number.
5. If you don't have online booking yet, start with our online booking and no-show reminders prompt.