Treat invented discounts as missing-slot hallucinations

You asked for a warmer follow-up. The draft offers ten percent off. You asked for a late note. It promises Friday morning. You asked for a proposal polish. It adds a complimentary inspection. None of those were yours.

The short answer

The sequence below is the control path: same vendor pages, same prompts, product names kept exact.

An invented discount, policy, or time is an empty slot, not a tone problem. Fill the fact, mark the gap [NEED], and desk-check every price and promise before you send. The same check works in ChatGPT, Claude, and Gemini.

Why the helpful sentence is the dangerous one

The draft is trying to complete the email. An empty price becomes a discount. An empty time becomes "soon" or "Friday morning." An empty scope becomes a free add-on. Anthropic documents ways to reduce hallucinations by grounding and by allowing "I don't know." OpenAI's prompt guide puts the constraints in before you generate. Gemini can also attach a source that does not support the sentence. Workspace sources help says it can miss a source, cite the wrong file, or present a source it made up.

Your version of "I don't know" is [NEED], plus a person who sends.

What the empty slot becomes

• Empty price → a discount or a "standard package"
• Empty policy → a flexible cancel or a free add-on
• Empty time → soon, this week, a priority slot
• Empty scope → an extra visit or a free consult

Fix the slot. Do not ask for a warmer tone until the facts are right. The prompt pattern is job, facts, and NEED. The email version is customer emails.

The desk check, in every tool

Before anything a customer will read leaves the building:

1. Would you say this at the desk?
2. Is every price, time, and promise one you already approved?
3. Is there exactly one ask?
4. Did the draft invent a policy, a discount, or a timeline?
5. Is any [NEED] still blank? Fill it or delete the line.

How each tool still invents, and the specific check

ChatGPT

Breadth is the risk. An image caption, a research snippet, and the email can land in one reply, and a snippet's price can leak into the email. Split the jobs. The email gets only the facts list. If web research is on, no researched price enters the customer draft until you have opened the page.

Claude

Long drafts invent in the middle, where you stop reading. Check the exclusions and the price table, not only the opening paragraph. Project instructions should say never invent a price or a policy. Instructions do not replace the desk check.

Gemini

A Sources chip is not the fact. If the chip is missing, that reply had no links. If the chip is present, open it. Inside Workspace, confirm whether web search is even on. A cited Drive file can be the wrong file. The rate sheet you meant is the one you open.

Try this on purpose

1. Ask for a "nicer" follow-up with no price and no ask.
2. Highlight every invented commitment.
3. Re-run with the job, the facts, and [NEED].
4. Confirm the inventions are gone.
5. Save both drafts in a folder named hallucination-examples so the next hire sees the pattern.

What not to send unsupervised

Refunds, medical or clinical claims, legal threats, and angry disputes. Drafting is fine. Sending those still needs a person. Regulated advice does not get more true because the paragraph is smooth.

Example (illustrative): a salon running a Tuesday special

Problem: the draft announced 15 percent off a service that was never on special. What was built: the price line deleted, the real service name provided from the book, one ask (reply to book). Tools it lives in: the same chat, second pass. Owner still does: the send. Payoff: not measured.

Mistakes and limits

• A warmer-tone request used as a fix for a wrong fact.
• Trusting the last paragraph of a long draft you did not read.
• A Sources link you did not open.
• Standing instructions with no desk check.
• Saving only the bad draft's apology, not the pair of examples.

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

Is this a bug in one product?

No. Any of the three will fill an empty slot. The check is the same. The place a false source shows up is different: a researched snippet in ChatGPT, a buried line in a long Claude draft, a Sources chip in Gemini.

Can a Project or a Gem stop this by itself?

It lowers the chance. It does not close it. The message still needs today's facts, and a person still sends.

What do I do with a [NEED] I cannot fill?

Delete the line or wait. Do not tell the tool to "use your best guess."

Should I turn web search off?

For a customer email, yes, unless you are ready to open every link before you send. For a vendor compare, search can gather candidates. Those candidates stay out of the customer email until you have checked them. That compare is its own article.

Keep these

• Empty slot, not bad tone.
• [NEED] for anything you did not type.
• Desk-check price, time, policy, and the single ask.
• Open a cited page before you trust it.

Sources

• Anthropic: reduce hallucinations
• OpenAI prompt engineering
• Sources in Gemini Apps
• Workspace sources

Related reading

• The prompt block
• Quote drafts that keep your prices
• Customer emails