"Washer 7 has been broken for two weeks. Again."
That review sat at the top of Bluebird Wash and Fold's (a made-up laundromat with a wash-and-fold pickup service) Google page. Marisol, the owner, wanted to know what customers were really unhappy about before she spent money fixing anything. So she asked an AI.
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 Use an Internal AI Tool to Analyze Customer Feedback," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by Wade Chambers, who demos Amplitude's internal AI tool, Moda (the How I AI pages don't state his title), in the episode "How I AI: Amplitude's Viral Internal AI Tool for Product Development." Read the original workflow on How I AI. Watch the full episode. Guest site: none listed on the How I AI workflow page; company site: Amplitude.
What the original does
Amplitude built a tool that searches support tickets, call notes and other company systems, then narrows from broad themes to the records behind them. The workflow runs in 4 steps:
- Aggregate the inputs from approved sources, keeping the source, date, account and a link on every record.
- Run a first theme pass that reports the sample before listing themes, with frequency, distinct accounts, exact quotes, counterexamples and confidence.
- Drill into one theme by customer job, segment, workflow stage and consequence, linking every claim to the original record.
- Pull the best quotes with their source and a one-line summary of each customer's problem.
The runbook names the trap we walked into: "A recent or convenient sample is presented as the full customer picture."
The test
Everything here is made up.
Over 90 days, Bluebird collected feedback in 3 places: 64 Google reviews, 118 customer texts and 23 comment cards. That is 205 pieces of feedback.
Run 1, the original as written: Marisol exported the Google reviews, because that was easy, and pasted them in with the first-pass prompt. The AI did what the prompt asks. It reported 1 source, 64 records and no exclusions. Then it listed themes:
- Machines out of order: 19 reviews from 12 reviewers. Top theme, high confidence.
- Pickup service praised: 9 reviews called pickup "always on time." Only 3 complained.
- Its advice: fix or replace machines first.
Marisol got a $2,400 quote for a new washer that week.
The report was honest about what it saw. It just couldn't see the texts. In a chat app, "accessible sources" means whatever you pasted, so "no exclusions" read like the full picture.
The rerun
Run 2: same feedback, our prompt below. It made Marisol list every place feedback lands before any analysis, so all 205 records went in, labeled by source. Names became initials and phone numbers became the last 4 digits. She tagged each customer as walk-in or pickup subscriber and added what subscribers pay: $38 a week in this example.
The picture flipped:
- Late pickups and deliveries: 31 records (27 texts, 3 reviews, 1 card) from 17 named customers plus 1 unsigned card. One subscriber sent 6 of those texts alone, and the prompt counted her once.
- Machines out of order: 27 records (19 reviews, 6 cards, 2 texts). Of those, 15 named the same machine, washer 7.
- Counterevidence kept: the 9 on-time reviews were real. Of the 27 late texts, 22 came on Saturdays, and 7 of the 9 happy reviewers used the weekday route.
Then it followed the money. Of the 17 named customers who complained about late pickups, 5 had canceled their weekly service. That is 5 x $38 = $190 a week, or $190 x 52 = $9,880 a year walking out over one Saturday route.
The machine theme got an honest answer: dollars "unknown," because walk-in reviewers can't be tied to spending. Its advice was to fix washer 7 and check whether that clears most of the complaints, not to buy a new machine.
Marisol added a second Saturday driver for 4 hours at $18 an hour, or $72 a week. That is less than the $190 a week the 5 lost subscribers had been paying.
The loudest channel is not the same as the costliest problem.
Verdict: 8/10
Strengths: One of the most careful prompts we've reviewed. It asks for the sample before the themes, counts distinct customers, keeps counterexamples, and bans invented quotes outright. The drill-down step, by workflow stage, is what found the Saturday pattern in Run 2.
Weakness: It was built for a company tool wired into permission-aware systems. A business owner using a chat app has to supply the sources by hand, and nothing in the prompt makes you list the ones you skipped. It also ranks themes by frequency, so a problem that 9 reviewers mention can outrank one that cost you 5 paying customers.
Founder tweak: Before any analysis, list every place feedback lands and paste them all, labeled. Tag customers by type and spend, count repeat complainers once, and rank themes by dollars at risk, not mentions.
Copy the MCAFax version
This is our own rewrite, not the original. Copy the original from Wade Chambers on How I AI.
I own [business type]. I want to find out what my customers are really unhappy (and happy) about, and which problems are costing me money.
Do not analyze anything yet.
PART 1. Inventory first.
Ask me to list every place customer feedback lands: online reviews, texts, emails, voicemails, comment cards, staff notes and cancellation reasons.
For each one, ask roughly how many items I have for [date range] and whether I can paste or export them.
Tell me which sources are still missing before we start, and remind me that you can only see what I paste.
PART 2. Clean the input.
Have me paste each source under its own label, with dates.
Replace customer names with initials and phone numbers with the last 4 digits. Never accept card numbers, home addresses or health details.
Tag each item with a customer ID (initials plus last 4), a customer type I define (for example [walk-in / subscriber]) and, if I give it, what that customer spends per [week / month].
PART 3. Themes.
Report record counts per source first.
List themes with: number of records, number of distinct customers, which sources they came from, 2 exact quotes with source and date, and any feedback that points the other way.
Count a customer who repeats the same complaint once. Never invent or reword a quote, and label any summary as a summary.
PART 4. Follow the money.
For each theme, show which customer types it affects, how many of those customers stopped buying or canceled, and the dollars at risk per week and per year, using only numbers I gave you. Show the math.
If you can't tie a theme to money, say "unknown" and tell me what I'd need to check.
PART 5. Drill down and decide.
For the top 2 themes by dollars at risk, break them down by day, location, staff member or service where the data allows.
Give me one fix to test for each, what it costs, and the 3 customers I should call first to ask what happened.
Paste into any AI chat, and read the quoted records yourself before you spend money on a fix.
Try this week
- List every place your customers leave feedback, including texts and staff notes.
- Export or copy 90 days from each one, and strip names down to initials.
- Tag each customer by type and, where you know it, weekly or monthly spend.
- Run the prompt above and rank themes by dollars at risk, not mentions.
- Keep the good reviews coming with our automated review requests prompt, and turn your top answers into a help page with the FAQ and help center builder prompt.
More owner prompts: browse the MCAFax prompt library.
