WSJ Prime7.00%▲ 25 bpsSOFR3.87%▲ 23 bpsFed Funds Target4.00%▲ 25 bps10-Yr Treasury5.22%▲ 42 bpsDiesel, retail$6.199▲ 60¢Bank C&I Lending$2,959B▲ 0.82%WSJ Prime7.00%▲ 25 bpsSOFR3.87%▲ 23 bpsFed Funds Target4.00%▲ 25 bps10-Yr Treasury5.22%▲ 42 bpsDiesel, retail$6.199▲ 60¢Bank C&I Lending$2,959B▲ 0.82%

Front Page › Markets & Main Street › Small Business

AI Prompts, Reviewed

My AI ghostwriter nailed my voice. Then it offered a $375 discount I never approved.

We trained AI on a sign shop owner's sent mail. It nailed his voice, then offered a $375 discount he never approved. Here's the fix.

An uncapped fountain pen and heavy shears slicing a paper invoice on a sign maker's workshop table.
AI Prompts, Reviewed · Markets & Main Street

"No rush at all! Happy to knock 10% off if you can pay this week."

A sign shop owner asked AI to chase a $3,750 invoice, 41 days late. The draft sounded exactly like him. That was the problem.

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 "Train Claude Cowork to Write Emails in Your Personal Style," a workflow from How I AI (hosted by Claire Vo, written up by the ChatPRD team), shown by JJ Englert of Tenex in the episode "JJ Englert's Guide to a 'Daily Operating System' with Claude Cowork." Original workflow: https://www.chatprd.ai/how-i-ai/workflows/train-claude-cowork-to-write-emails-in-your-personal-style . Full episode: https://www.chatprd.ai/how-i-ai/jj-englerts-guide-to-a-daily-operating-system-with-claude-cowork

What the original does

JJ connects Gmail to Claude Cowork and asks it to study the emails he sent in the last 30 days. Cowork turns his tone and structure into a reusable writing skill, so every future draft sounds like him. Then he adds one rule we love: never send anything, only draft.

The test

Composite (made-up) data and simulated drafts for a sign shop: the owner plus 2 installers. We gave the AI the owner's last 30 days of sent mail, 84 emails:

  • Quote replies to new customers: 52 (62%)
  • Supplier and vendor emails: 19 (23%)
  • Install scheduling: 9 (11%)
  • Collections: 4 (5%)

That's 52 + 19 + 9 + 4 = 84.

The job: a polite but firm reminder for invoice #1187. 15 storefront signs at $250 each, so 15 x $250 = $3,750. Due August 29. Sent October 9, which makes it 41 days late.

The skill summed him up as "warm, upbeat, closes with an offer." Two problems:

  1. It learned one kind of email as the whole voice. Most of the sample was quote replies, so the collections email came out sounding like a sales pitch.
  2. It made up a promise. That "10% off" is $375. The customer would have paid $3,375. The owner never offered a discount to anyone.

The page's own notes flag this: "A draft invents a commitment or personal detail". The short prompt doesn't stop it, so we rewrote it.

The rerun

We pulled 90 days instead of 30: 241 sent emails. Of those, 11 were collections. We had the AI split traits that never change (short sentences, signs off with first name, no jargon) from traits that depend on the email type.

The new draft:

"Hi Dana, invoice #1187 for $3,750 was due August 29 and is now 41 days past due. Can you send payment by Friday, or tell me what's holding it up? [CONFIRM: late fee per our contract?] Thanks, Sam."

Same voice. No discount. And a flag where the AI didn't know the terms instead of a guess.

A good voice copy sounds like you. A good assistant knows which you.

Verdict: 8/10

Strengths: it's the easiest setup we've reviewed, since you teach the AI from emails you already wrote. The draft-only rule is the best safety habit in the whole library.

Weakness: 30 days is one season of your inbox. A busy quoting month teaches your sales voice. Then it writes your hardest emails in it.

Founder tweak: pull at least 10 sent emails of each type you write: quotes, vendors, scheduling, collections. Make the AI mark anything it can't prove from your brief with [CONFIRM].

Copy the MCAFax version

This is our own rewrite, not the original prompt. Copy the original from JJ Englert on How I AI: https://www.chatprd.ai/how-i-ai/workflows/train-claude-cowork-to-write-emails-in-your-personal-style

You are my email ghostwriter for my small business.
I will paste or connect sent emails from the last 90 days, sorted by type: quotes, vendors, scheduling, collections, and customer complaints.
Step 1. Count the emails of each type. If any type has fewer than 5, tell me before you go on.
Step 2. List the voice traits that show up in every type: sentence length, greeting, sign-off, words I use, words I never use.
Step 3. For each type, list what changes: how direct I am, how long, how I ask for something.
Step 4. Save this as my voice guide, with one short sample email per type.
Step 5. When I give you a brief, write the draft in the voice for that email type.
Rules: Never send anything. Drafts only.
Never offer a discount, deadline, refund, or promise unless it is in my brief.
Mark any fact you can't find in my brief as [CONFIRM: ...].
Never copy a sentence from an old email word for word.
My brief: [example: invoice #1187, $3,750, due Aug 29, ask for payment by Friday]

Paste into Claude, ChatGPT, Gemini, or Copilot.

Before you hand off your inbox

  1. Pull 90 days of sent mail, not 30.
  2. Sort it by type and count each pile.
  3. Add 10 or more of your toughest emails, such as late payments.
  4. Run the MCAFax version and read the voice guide line by line.
  5. Test it on one collections email. Check every number and promise.
  6. Keep your reminders on schedule: https://mcafax.com/notebook/prompts/invoice-payment-reminders (more prompts at https://mcafax.com/notebook/prompts)

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

More from AI Prompts, Reviewed

The series →

The AI built my plan around one $4,800 job. 650 hours of my calendar said no.

Small Business · Oct 9

The AI called my call "warm and professional." I lost the $1,860 job.

Small Business · Oct 8

The AI turned one $45 favor into a $4,680 policy.

Small Business · Oct 7

More in Markets & Main Street

All →

Fervo Energy (FRVO): A Short Seller Challenges the Geothermal Data Ahead of a November 10 Lock-Up Expiry

From Wall Street · Oct 9

Snap (SNAP): Two Activist Funds, a $5 Billion Specs Price Tag and a Contrarian Case for 50% Upside

From Wall Street · Oct 9

LuxExperience (LUXE): Can the Net-a-Porter Turnaround Push the Stock Out of Its $10 Range?

From Wall Street · Oct 9

Getting the calls? Shut it down.

MCAFax was built to do the thing you can't do from inside the room: make the phone stop. Check any broker against our member-built database, send cease & desist letters from your own Gmail, and build the paper trail. Some laws, like the TCPA, put statutory damages on illegal calls — whether they apply to your calls depends on your situation, and business lines get less protection than home ones. It's free, and we're not a law firm.

Join the network