Dialing In the Tone of an Answer Without Spending an Hour

A hairdresser has received a two-star review: "Forty-five minutes of waiting despite my appointment, and nobody apologized." She asks the AI for a reply that is "professional and friendly." The text that comes back begins with "We are truly sorry for this disappointing experience," and she recognizes everything in it except her own way of speaking. She rewrites it by hand, which takes her longer than if she had done it all herself.

Tone is hard to ask for because adjectives aren't instructions. "Friendly" covers the buddy, the salesperson, and the switchboard operator. Here are three more reliable ways to dial it in.

Situate the reader and the relationship

Tone follows from who is talking to whom, in what situation. Say that, instead of naming a mood.

Write the public reply to this customer review: "[PASTE THE REVIEW]".

I run my hair salon alone in [CITY]. I answer the way I speak to a customer I see often: polite but personal, short sentences, no stock phrases. I acknowledge the wait, explain in one sentence what happened ([WHAT HAPPENED]), and offer for her to call me so I can sort it out. 60 words maximum.

"The way I speak to a customer I see often" does more work than five adjectives, because the AI knows what that kind of sentence looks like.

Show a model

When you've already written a text that sounds right, paste it and ask for the description before asking for the imitation.

Here is a message I wrote whose tone I like:

"""
[PASTE YOUR MESSAGE]
"""

Describe this tone in five concrete points: sentence length, level of language, what I say and don't say, how I open and close. No vague adjectives like "warm."

You get a sheet you can paste at the beginning of your future prompts. If tone matters a lot to you, for example across all your communications, the article on getting AI to write in your style goes further. For one-off texts, two pasted examples are enough, as few-shot prompting explains.

Move the slider one notch

You don't always have a model, but you know how to recognize too much or not enough. Rather than starting over, shift the existing text:

• "Same text, one notch more direct."
• "Same text, removing everything that sounds like corporate apologies."
• "Same text, as if I were writing to someone I'm on a first-name basis with."
• "Same text, but without the salesperson tone: no superlatives, no exclamation marks."

The first two words, "same text," keep the AI from reinventing the substance while it changes the form. Without them, it may change the facts without warning.

If you're torn between two settings, ask for both side by side: "Give two versions of the same message, one very restrained and one warmer." You choose, then you can refine.

The pitfalls to know

The "polite" tone that turns servile. AIs readily add apologies and thank-yous by the barrage. If your reply must stay firm, for example to refuse a refund, say so: "polite but without apologizing for what isn't our fault."

The register of language. Formal or casual (first names and contractions, or titles and full forms) is a decision to make yourself and to write down. Without an instruction, you'll get one or the other at random, sometimes both in the same message.

Facade emotions. When the AI writes "we understand your frustration," it's a formula you could find in a thousand customer service replies. If the message touches real suffering, a death, a layoff, write it yourself or rewrite what it suggests in your own words. When not to use AI returns to these cases.

To try with a real message

Take an email you have to send this week that you find difficult: a refusal, a follow-up, a reply to criticism. Ask for a first version using the reader method, then move the slider one notch. Read it aloud before sending: if a sentence bothers you when you say it, it will bother the recipient.