How to soften a message: 5 moves that keep the ask, and the AI prompt that makes all five

How do you soften a message without losing the ask?

You soften a message by changing its delivery and leaving its content alone: every fact, request and deadline stays, and every jab goes. Five moves do almost all of it. Here they are on one real draft, a note to a colleague called Dan whose deck had wrong numbers for the third time.

The last row is where most softening goes wrong. "A bit" and "slightly" are fine on a judgement about the work; the moment they reach the request ("could you possibly, whenever you get a chance"), the deadline is gone. Softening should cost you adjectives, never nouns.

The reason to bother at all is measured. Kruger, Epley, Parker and Ng ran 154 pairs of students, friends and strangers, through a 2005 study in which each sent statements meant to be sarcastic, serious, angry or sad. Senders predicted their email tone would be read correctly 89.3% of the time. Receivers got 62.8%. The same senders talking face to face predicted 89.4% and were read at 73.9%.

In the paper's second study, email readers spotted sarcasm at "an accuracy rate indistinguishable from chance (50%)"; in the third, friends were as overconfident as strangers. A reader who can't tell your feeling from your request fills the gap with the worst reading, which is why the fix is clarity about the request more than warmth.

Which built-in tone buttons can be told what to keep?

Only the typed instruction boxes can. As of 26 September 2026, the one-tap presets in Gmail (Formalize, Friendly, Shorten) and Apple Writing Tools (Friendly, Professional, Concise) take no instructions, so none of them can be told which deadline to protect; the free-text box beside each one can. I compiled this from each vendor's own help page on that date.

Google's own help page makes the point for me. Its Help me write instructions give "Mention the Friday deadline" as an example of what to type into the prompt bar, which is the must-keep list in four words. The Friendly button next to it has no way to know the Friday deadline exists. Apple's Writing Tools guide and Microsoft's coaching page split the same way: Apple pairs its three presets with a describe box, and Microsoft pairs Coaching, which only suggests fixes, with a Draft with Copilot prompt box that takes your own instructions.

The decision rule: use a preset for a two-line reply with nothing at stake, and the typed box, or the prompt below, for anything with a number, a date or a request in it.

The ChatGPT prompt that softens a message and keeps every deadline

Paste your draft into ChatGPT, Claude or Gemini with this above it. The bracketed lines are yours to fill; the last three lines are the ones that stop the rewrite turning into mush.

Here is a message I drafted. I am annoyed and it shows. Rewrite it so it lands without sounding rude or passive aggressive.

Who it is to: [my colleague Dan, same level as me, we work together every day]

What I still need to say: [the numbers were wrong three times; I need slides 6 and 9 fixed by 9am tomorrow; I need him to check the totals before he sends]

What I want to drop: [the 11pm complaint, anything that sounds like I am his manager]

Keep it under [100] words, about the length of my draft. Keep my wording wherever it is already fine. Turn orders into requests but keep every date exact. Do not add apologies, compliments I did not write, hedges or "hope you are well". Do not soften the ask itself, only the way it is said.

After the rewrite, list in one line each what you changed and why.

My draft: [paste]

Each line does one job. The relationship line sets the register, because the same sentence is fine to a friend and a firing offence to a client. The "still need to say" line is the list the model is not allowed to lose. The length cap stops the padding that makes a softened message read as evasive. "Keep my wording" protects the sentences that were never the problem, and the change list at the end is how you spot a dropped deadline in five seconds.

The same prompt works in Gmail's prompt bar or Apple's describe box if you shorten it to the must-keep list and the no-apologies line. For writing an email from nothing rather than editing one, the email drafting prompt is the better start; for a message whose whole job is refusing something, use the decline templates.

Before and after: what a good rewrite keeps and what mush drops

A good softened message is about as long as the draft and carries every noun the draft did. I wrote all three versions below myself as a worked example of the prompt above, not as a measurement of any model.

The draft, as typed at 11pm. "Dan, this is the third time the numbers in the deck were wrong. I had to redo the whole thing at 11pm again. I need you to actually check your work before you send it to me. Fix slides 6 and 9 by tomorrow morning."

The rewrite the prompt should produce. "Dan, slides 6 and 9 still have the wrong Q3 numbers, and this is the third deck in a row I have had to rebuild the night before. I need a version I do not have to check. Can you fix both slides and run the totals against the source sheet before you send it, by 9am tomorrow?"

The rewrite you will sometimes get instead. "Hi Dan! Hope you are doing well. I just wanted to flag a small thing: I noticed a couple of the numbers on slides 6 and 9 might need another look. Totally understand things get busy! Whenever you get a chance, could you possibly take a peek? Thanks so much for all your hard work on this."

The good rewrite is one word longer than the mush and carries three things the mush lost: the pattern, the deadline and the process ask. If a rewrite could have been sent by someone who wasn't annoyed, it lost something. Put it back.

How far should you soften? Stop at slightly warm

Slightly warm is the stopping point. Boomerang read the reply rates on 40 million emails in 2016 and found that replies "peaked about 15% higher than neutral for emails with a slightly warm tone", then fell as the praise piled up; a gushing email scored above 0.9 on its sentiment scale and did about as well as a neutral one. For a complaint, a moderately negative email beat the same content in a neutral tone by 13%, and poisonously negative ones were the least likely to get any reply.

So you do not have to scrub the feeling out. A message that sounds mildly annoyed and asks for one clear thing gets more replies than one that sounds like nothing happened. One caution on the data: Boomerang counted replies, not outcomes, and of its most opinionated emails it wrote that it had no idea whether they drew positive responses or declarations of war.

The two ways an AI rewrite goes wrong, and the line that fixes each

It softens into mush. A model reads "make this sound better" as "make this sound like nothing happened". Deadlines become "whenever you get a chance", facts become "I noticed", and the ask acquires a "possibly". This is the default failure, and it's worse than the blunt draft because the reader now has to guess whether you meant it. If it happens, reply with one correction: "Too soft. Put the 9am back and cut every hedge: just, might, possibly, a little, whenever."

It adds apologies you did not mean. "Sorry to bother you", "I totally understand", an exclamation mark, a compliment to balance the complaint. None were in your draft, and each tells the reader you're less sure than you are. The no-apologies line in the prompt removes most of them; the change list catches the rest, because an added apology shows up as a change you didn't ask for.

A smaller third failure is the model rewriting sentences that were fine, which is what makes a softened message sound like software. Swap one phrase back to how you would say it out loud before you send. Running the same prompt in a second model and keeping whichever version held the 9am is a quick cross-check; free ChatGPT and a free Claude or Gemini account do it with a copy and paste, and Whizi's compare view does it in one window. The empty chat guide has more correction lines of this kind.