Practical Prompt

Two rules that stop AI from making things up in customer emails

One catches what's missing, the other proves what's there. Together they cover most of the ways AI goes wrong with customers.

Written by Tauri on 7 October 2026. About 5 minutes to read.

The worst AI mistake with customers isn't a clumsy sentence. It's a confident, friendly, completely made-up policy. "Of course, you can return it within 60 days!" when your limit is 14. The reply sounds perfect, which is exactly why it slips through.

I've used one rule against this for a while. A reader on Reddit suggested a second one, and the two fit together so well that I'm adding it to my routine.

Rule 1: make it flag the gaps

If the AI doesn't know your policy, it won't say so. It will guess something that sounds reasonable. So tell it what to do instead:

The gap rule
Only use the facts I've given you. If the answer isn't in my facts, write [CHECK] instead of guessing.

Get one tested prompt like this every Tuesday. Sign up below, it’s free.

Now every gap shows up in the draft as [CHECK], and its biggest weakness turns into a short to-do list for you.

Rule 2: make it show where each fact came from

The gap rule catches what's missing. It doesn't prove that the rest is right. That's where the reader's idea comes in: ask the AI to point to the exact line in your facts behind every claim it makes. No source line means it's guessing.

This works best if you number your facts, so checking takes seconds.

The full prompt

Reply with sources
You help me reply to customer emails for [BUSINESS NAME].

My facts:
1. [e.g. Orders ship within 2 working days]
2. [e.g. Returns accepted within 14 days, unused items only]
3. [ADD YOUR OWN, one per line]

Customer email:
[PASTE EMAIL]

What I want: [ONE LINE, e.g. explain the return limit kindly, offer store credit]

Rules:
- Write directly to the customer, as the email they will receive.
- Only use the facts above. If the answer isn't there, write [CHECK] instead of guessing.
- The facts rule applies only to factual statements: policies, prices, reasons, dates and next steps. Greetings, thanks and friendly wording don't need sources.
- If you need information from the customer, ask them in the email. Use [CHECK] only for things I need to look up myself.
- After the reply, add a line starting with "Sources:" that lists which fact number backs each factual statement, quoting the exact line.
- Anything without a source counts as a guess. Remove it or mark it [CHECK].

The first, third and fourth rules weren't in my first version. I added them after testing it on a real email, which I'll get to below.

Then read the Sources line before the reply. If a statement has no fact number next to it, or the quoted line doesn't actually say what the reply claims, fix it before anything goes out.

What happened when I tried it

I tried the first version of this prompt on a real customer email: a question about a charge on their invoice. Here's how it went.

The gaps did their job. I started with only two facts. The first draft came back with [CHECK] where information was missing, and that told me exactly what to add. It took three more facts before the reply had everything it needed. Every time I added one, the Sources line updated, so I could see which part of the reply each new fact had changed.

Problem one: it wrote about the customer, not to them. The draft read like an internal note, the kind you'd send a colleague to explain the situation. I had to tell it to write directly to the customer and to ask them a question about what they'd done, since that was something only they could answer.

Problem two: the rules were too strict. When I asked for a friendlier tone, it put the greeting and the thank-you inside [CHECK] marks, because they weren't in my facts. It was treating "thanks for getting in touch" as a factual claim. The rule was working a little too well.

That's where the three new rules in the prompt come from: write to the customer, don't demand sources for politeness, and ask the customer directly when only they know the answer.

To be upfront: that was one email. I haven't used the fixed version on many more yet, so I'll update this note once I have.

Right facts can still sound like a template

Even with every fact correct, a reply can read like it came from a help desk bot. The reader pointed this out too, and my own test proved it. What helps me most is pasting two or three replies I've written myself as examples, which I explain in the customer email routine.

Before you hit send

  • Every [CHECK] is answered or removed.
  • Every factual statement has a source line that really says what the reply claims.
  • The Sources line is deleted from the final email.
  • You'd be comfortable if the customer screenshotted it.

Thanks to the Reddit reader who suggested the source line. If you have a rule of your own that keeps AI honest, write to me. The good ones end up here.

A note in your inbox on Tuesdays

Once a week I send one task, the prompt I use for it, and anything I got wrong along the way. It takes two minutes to read.

Send me the notes

Free. Unsubscribe whenever you like.