Thistlewick Dental never said $18,720. The case study did.

"We save about two hours a week of front-desk mess."

That's what the office manager said on a short call. The AI's draft put $18,720 a year in bold and wrapped a cleaned-up paraphrase in quote marks. The customer never named a dollar.

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 Automate Case Study Creation with an AI Note-Taker and a Custom GPT," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by Luke Harries, Head of Growth at ElevenLabs, in the episode "How I AI: 3 AI Workflows to Save $140K & Automate Marketing with ElevenLabs' Luke Harries" (June 2, 2025). Original workflow: https://www.chatprd.ai/how-i-ai/workflows/how-to-automate-case-study-creation-with-an-ai-note-taker-and-a-custom-gpt . Full episode: https://www.chatprd.ai/how-i-ai/ai-workflows-to-automate-marketing-with-elevenlabs-luke-harries

What the original does

Luke records a short customer interview in Granola, asks for concrete use and impact, then feeds a custom ElevenLabs copy-editor GPT both the summary and the raw transcript. The GPT drafts a case study and a social post in company voice. He tightens the instructions when the same errors keep showing up. Tools on the page: ChatGPT, Zapier, Granola. Optional Zapier comes after a human still reviews and approves.

The test

Composite (made-up) test data and simulated AI outputs. Every business, name, quote, and figure below is invented, and the drafts were simulated to illustrate failure modes the workflow page itself warns about. They are not a live Granola, ChatGPT, or Zapier run.

The setup: a commercial cleaning company, owner plus 4 staff. Night-cleaning contract with Thistlewick Dental (a made-up practice): $1,250/month, 12 x $1,250 = $15,000/year. Clock: Monday, Oct 12, 2026. The manager consented to a short interview about switching to nights.

We simulated it the way most owners would run it: paste the Granola summary into a custom GPT with brand voice and a case-study outline. No source table, no [VERIFY] rule, no customer-approval gate.

1. Invented metric. The transcript said "about two hours a week." The simulated draft claimed Thistlewick Dental "saves $18,720 a year." Implied math: treat those hours as clinical chair time at $180/hour, 2 x $180 x 52 = $18,720. The customer never said $180, never said chair time, and never approved a dollar figure. Honest front-desk math at $18/hour: 2 x $18 x 52 = $1,872/year, and even that is an estimate, not a quote.
2. Paraphrase in quote marks. The manager said the waiting room used to be a mess when they opened. The simulated draft quoted: "Our waiting room used to be a disaster every morning. Night cleaning fixed that overnight."

Ship that without sign-off and you risk the $15,000 account. The page's own rule: "Never invent metrics, customer intent, product capabilities, pricing, or links."

The rerun

We forced a source table before any prose, a [VERIFY] tag on every number and quote, and a hard stop: no publish and no social until the customer approves final text, attribution, and channels. Social may only reuse lines from the approved case study.

Simulated result: the draft keeps "about two hours a week," shows $1,872 as [VERIFY: wage x hours], drops $18,720, and leaves the waiting-room line as paraphrase. Nothing posts until Thistlewick Dental replies yes.

Exact quotes only. Invented dollars don't belong in quote marks.

Verdict: 8/10

Strengths: summary plus raw transcript, a voice-locked copy editor, and the ban on invented metrics. Social only from an approved case study stops a second round of fiction.

Weakness: Granola + custom GPT + Zapier is a lot for a 5-person shop. Without the source table and approval gate, the workflow still ships a polished lie.

Founder tweak: force a source table (quote / paraphrase / metric / [VERIFY]) before any paragraph, and block publish until the customer signs off on text, name, and channels.

Copy the MCAFax version

This is our own rewrite, not the original prompt. Copy the original from Luke Harries on How I AI: https://www.chatprd.ai/how-i-ai/workflows/how-to-automate-case-study-creation-with-an-ai-note-taker-and-a-custom-gpt

You are my case-study copy editor for my small commercial cleaning company.
I will paste: (A) interview summary, (B) raw transcript, (C) product facts I approve, (D) banned claims.
Step 1. Build a SOURCE TABLE with one row per claim:
- Type: Quote | Paraphrase | Metric | Product fact
- Exact words from transcript (or "not in transcript")
- Proposed wording
- Status: OK | [VERIFY: what is missing] | DROP
Step 2. Only after the table, draft the case study: problem, what we did, result, short customer line.
Step 3. Draft one social post using ONLY lines marked OK in the table and already in the case study.
Rules:
Never invent metrics, pricing, savings, product capabilities, or links.
Never put paraphrase in quotation marks.
Every $ figure must show the math and a source row. If the customer did not state a $, mark [VERIFY] or DROP.
Banned claims: [list yours].
Do not say this is ready to publish. End with: CUSTOMER APPROVAL REQUIRED for final text, name/attribution, and channels.
My inputs: [paste summary + transcript]

Paste into ChatGPT (Custom GPT), Claude, Gemini, or Copilot.

Before you publish a customer story

1. Get consent in writing for the interview, quote style, and channels.
2. Keep the raw transcript. Never draft from the summary alone.
3. Run the MCAFax version and read every SOURCE TABLE row before the prose.
4. Send the customer the final text. No website, no social, until they approve.
5. If a metric needs a wage they never gave, mark [VERIFY] or cut it.
6. More prompts: https://mcafax.com/notebook/prompts

SBA on managing your finances: https://www.sba.gov/business-guide/manage-your-business/manage-your-finances