The AI watchdog flagged $38 in napkins. It missed a $21,320 cancellation.

The AI watchdog flagged $38 in napkins. It missed a $21,320 cancellation.

"Top problem: Pemberfold Diner, missing napkin bundle, 4 emails. Urgent."

That was #1 on an AI health check for Larkspindle Linen (a made-up commercial laundry). The napkins were worth $38. The report never mentioned that Hollowmere Cafe (made up) had emailed a 30-day cancellation. Hollowmere pays $410 a week. That is $410 x 52 = $21,320 a year walking out the door.

Rank problems by dollars, not by noise.

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 "Automate Business Health Checks with a 'Watchdog' AI Agent Playbook," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by Ryan Carson, solo founder of the B2B startup Untangle, in the episode "How Ryan Carson Manages 40 PRs a Day with Devin and Codex." Original workflow: https://www.chatprd.ai/how-i-ai/workflows/automate-business-health-checks-with-a-watchdog-ai-agent-playbook. Full episode: https://www.chatprd.ai/how-i-ai/how-ryan-carson-manages-40-prs-a-day-with-devin-and-codex. Guest site: none listed on the How I AI workflow page.

What the original does

Ryan runs a reusable playbook called Watchdog in Devin, an AI software engineer. When he wants a pulse check, the agent goes into each customer account, looks at activity since the last check-in (errors, bugs, anything odd), and returns the top 3 problems across all customers. Then it checks whether each one is fixed, in progress, or waiting on him. The page puts the goal in one line: surface the top three problems "avoiding raw data dumps."

The test

Everything here is made up. Larkspindle is in San Antonio: owner Teo, 2 vans, 18 restaurant accounts on weekly pickup, card autopay every Monday. Last check: Wednesday, Oct 7, 2026. Today: Sunday, Oct 18, 2026. That is 11 days.

Naive run (email and help inbox connected, "check every account since last time, top 3"):

1. Wrong window. With no saved date, the agent used the last 7 days. Hollowmere's cancellation came Friday, Oct 9. It never showed up.
2. Loud beat big. Pemberfold sent 4 emails about $38 of napkins. It ranked #1 because it had the most messages.
3. Missing source. Brindlesea Grill (made up) pays $1,240 a week. Its card was declined Monday, Oct 12, and the retry failed Thursday, Oct 15. The payment report wasn't connected. If the card isn't fixed by Monday, Oct 19, Teo is owed $1,240 x 2 = $2,480.
4. Closed is not fixed. Teo promised Quenby Tacos (made up) a $96 credit for stained aprons and closed the ticket Tuesday, Oct 13. The agent marked it "Fixed." The credit never hit the invoice.

The rerun

Same accounts, plus 4 rules: read the last-check date from a note and start there, list every source and say which ones weren't checked, rank by money at stake, and only call something fixed if a record proves it.

New top 3:

1. Hollowmere: cancellation effective Sunday, Nov 8. $21,320 a year. OPEN, no reply sent.
2. Brindlesea: $1,240 unpaid, $2,480 by Monday. OPEN, no new card on file.
3. Quenby: $96 credit promised, not on the invoice. NOT FIXED.

Napkins moved to "also seen." The unconnected route log was listed as NOT CHECKED, not guessed.

Verdict: 7/10

Strengths: top 3 instead of a data dump is the right shape for a busy owner. The status check keeps old problems from hiding. Reusable every week.

Weakness: built for a software founder with Devin and an error tracker. The page doesn't say how to rank, where the last-check date lives, or what proves a fix. No prompt to copy, only steps.

Founder tweak: save the last-check date, connect your payment report first, rank by dollars at stake, and make the AI show the record behind every "fixed."

Copy the MCAFax version

This is our own rewrite, not the original. Copy the original from Ryan Carson on How I AI: https://www.chatprd.ai/how-i-ai/workflows/automate-business-health-checks-with-a-watchdog-ai-agent-playbook

You are the weekly health checker for [business name].

You read and report. You never email, refund, or charge anyone.

Inputs I will paste or connect:

(A) date of my last check: [date]

(B) customer list with weekly or monthly value for each

(C) sources: email, help inbox, payment report, schedule

(D) anything I promised a customer, with the date

Step 1. Window = from (A) to today. Say the dates.

List each source in (C) as CHECKED or NOT CHECKED.

Step 2. For every customer in (B), find: failed or late payments,

Cancellations or "we're thinking of leaving," complaints,

Missed service, promises in (D) not yet done.

Step 3. Rank by money at stake: amount owed now, plus the

Yearly value of any customer who may leave.

Show the math. Message count does not raise rank.

Step 4. Top 3 only. For each: customer, problem, dollars,

Date first seen, and status: FIXED (name the record),

IN PROGRESS, or OPEN. A closed ticket is not proof.

End with: "also seen" (one line each) and today's date

So I can paste it into (A) next time.

Paste into Claude, ChatGPT, or Gemini with your files attached or tools connected.

Before you trust it

1. Write the last-check date somewhere the AI can read every time.
2. Connect your payment report. Failed payments are the quiet ones.
3. Give each customer a dollar value, even a rough one.
4. Open the record behind every FIXED. If there isn't one, it's OPEN.
5. Chase unpaid invoices on a schedule: https://mcafax.com/notebook/prompts/invoice-payment-reminders

More owner prompts: https://mcafax.com/notebook/prompts

FAQ

What do I need for the original workflow? Devin, a reusable playbook named Watchdog, and access to each customer account and its error and activity data.

Can I run this without Devin? Yes. Any AI assistant that can read your email, help inbox, and payment report can run our version.

Why did the AI miss a cancellation? It didn't know when you last checked, so it guessed 7 days. Save the date and make it start there.

How should the AI rank problems? By money at stake: what's owed now plus the yearly value of a customer who may leave. Not by email count.

When is a problem really fixed? When a record proves it: a cleared payment, a credit on the invoice, a completed job. A closed ticket is not enough.

How often should I run it? Weekly, the day before your billing day, so failed payments show up in time.