Write prompts that stop invented discounts

You asked for a warmer follow-up. The draft offers ten percent off. You never ran a discount. The tone was not the bug. The price slot was empty, so the draft filled it.

The short answer

Name the job in one sentence, list only the facts the draft must not invent, and mark gaps [NEED]. Then run a second pass that cuts filler and keeps every fact. The same block works in ChatGPT, Claude, and Gemini. Style words like "be professional" do not hold a price in place.

Why "be professional" fails

With no job and no facts, the draft fills holes with kindness: a discount, a timeline, a policy you never approved. OpenAI's prompt guide puts constraints and the reference material in the prompt before you generate. Anthropic's notes on reducing hallucinations say the same thing in their words: ground the draft, and let it admit it does not know. Your shop's version of "I don't know" is [NEED].

The block to paste

Voice: clear, calm, local. No hype. No invented policy.
Job: [one sentence]
Facts:
- ...
- ...
Draft [email / section / checklist]. Do not invent. Mark gaps [NEED].

Then, in the same thread: Tighten under 120 words. Keep every fact. One ask.

Job examples that fit in one sentence: confirm Tuesday 3pm and ask them to unlock the side gate. Follow up on Monday's quote with one yes or no. Draft the exclusions section from the rate rules you paste under Facts.

Where to put the block in each tool

ChatGPT

Paste the voice lines into the Project instructions or the custom GPT, and put today's job and facts in the message. If you have no system field, the block goes at the top of the message. See Projects and GPTs.

Claude

Put the voice and the never-invent lines in the Project instructions. Today's facts stay in the user message. A Project is the frozen prior. The message is the fresh job.

Gemini

Put the voice lines in the Gem instructions. Add the rate sheet under Knowledge only if that file is approved for that account. Today's scope stays in the prompt. Previewing a Gem does not save it. Click Save.

What not to dump in

Full threads, card data, charts, and unused context invite guessing. Extra material is not more helpful. It is more places to invent from. Sensitive rules live in the data article.

Try this on a follow-up you owe today

1. Write the job sentence on paper first.
2. List five facts at most.
3. Run the block.
4. Fill every [NEED] yourself.
5. Run the tighten pass.
6. Read it out loud. Would you say it at the desk? Is there a discount you did not approve?

If the follow-up you owe is a policy renewal, here's an insurance renewal reminder email template, with when to send each one.

Example (illustrative): a quote that went quiet

Problem: Monday's quote has had no reply. What was built: one job sentence ("ask for a yes or no on the Monday quote"), five facts, no new price. Tools it lives in: whichever tab already has the quotes workspace. Owner still does: the send. Payoff: not measured. The working version of this email job is drafting customer emails.

Mistakes and limits

• Asking for "more human" before the facts are right.
• A voice essay with no job sentence.
• Pasting the whole thread so the model can "get context."
• Leaving a [NEED] in the message you send.
• One prompt that tries to be the SOP, the quote, and the review reply.

Doing this inside the tools you already use

If you would rather have this exact job drafted from the software you already use, with you still approving every send, see how a custom build works.

FAQ

Do the three tools need different prompts?

The block is the same. The place you save the voice lines changes: Project or GPT, Claude Project, or Gem. Today's facts always stay in the message.

What does [NEED] mean?

A slot you did not fill. The draft must show it, not guess it. You fill it or you delete the line before you send.

How long should the facts list be?

Short enough that every line is true. Five lines is plenty for a follow-up. A quote narrative can be longer if every line is an approved rate or a scope bullet from the call.

Can I reuse a prompt that worked last time?

Yes. Keep the voice lines. Replace the facts. A saved prompt with last month's price still in it will send last month's price.

Keep these

• Job, facts, [NEED], then a tighten pass.
• Voice lines live in the saved workspace. Facts live in the message.
• One ask.
• Read it out loud before it sends.

Sources

• OpenAI prompt engineering
• Anthropic: reduce hallucinations
• Gemini custom Gem tips

Related reading

• When the draft invents a discount anyway
• Draft the quote from approved prices
• Customer emails, facts first