How to Tell an AI Exactly What Tone to Use in an Email

Most AI email drafts miss the tone not because the model cannot write, but because it received no tone spec at all. Ask it to write a professional email and you get something technically competent — measured vocabulary, logical structure, complete sentences — that does not sound like you and does not quite fit the relationship. The model filled in every gap you left open with its defaults, and its defaults are generic.

This guide gives you a five-step method for telling an AI what tone to use before you write the first word, not after you are already editing a draft that missed the mark. It covers the four building blocks of tone, a step-by-step procedure for specifying each one precisely, ready-to-use tone-spec blocks for eight situations you write most often, how the major platforms handle tone instructions differently, and what to do when a draft still does not land. The method works in ChatGPT, Claude, Gemini, and Copilot in Outlook.

Before you start: what tone is actually made of#

Tone in an email is not a single dial — it has four building blocks. Knowing them lets you describe what you want precisely instead of waving at make it warmer and hoping the model interprets that correctly.

Register is the formality level: formal, semi-formal, casual, or intimate. It shapes vocabulary choices, sentence length, and whether you open with the task or with acknowledgment of the person.

Warmth is how much care for the other person shows through the writing — the difference between Attached is the report and Hope your week is going well, I have attached the report. A high-warmth email puts the person before the task.

Directness is how quickly and confidently the ask arrives. In a high-directness email, the request is in the first line, stated plainly. In a lower-directness email, context arrives first and the ask follows once the reader has been prepared for it.

Hedging is the cushioning language — just, maybe, if that works for you, sorry to bother you — that softens a request. Small amounts read as polite. Too much buries the ask and signals that you are not sure whether you should be making it.

A model asked only to be professional typically sets register to semi-formal, warmth to low-medium, directness to medium, and hedging to light. That is not wrong — it is a reasonable guess at what you meant. Specifying all four inputs removes the guesswork and gets the first draft where you want it.

How to specify tone before you draft: five steps#

1. 1

   Name the register and add a comparison

   Pick one of four registers: formal (board member, regulator, new external contact you have not met), semi-formal (client you know, professional acquaintance), casual (coworker you like, recurring vendor), or intimate (close friend in a professional context). Add a comparison alongside the label — semi-formal, like emailing a client I have worked with for a year — because that calibrates the model to a relationship rather than to an abstraction. A bare adjective leaves it guessing at the edges.

2. 2

   Paste two sample sentences you would actually write

   Find two lines from your own emails — from your sent folder, from a draft you liked, from anything you wrote that sounded right. They do not need to be about this topic. Paste them in and label them: Here are two lines that match the tone I want. The model calibrates to your register and rhythm without you needing to describe your voice abstractly.

3. 3

   Add two anti-examples

   Name two things you do not want: a phrase, a style, a tone register. Examples: do not write like a press release, do not open with I hope this email finds you well, avoid passive constructions. Anti-examples are often more useful than positive examples because they define the edges of what fits and give the model a clear correction signal even when your desired tone is hard to articulate.

4. 4

   Set warmth and directness as separate numbers

   State each on a scale of one to five. Warmth 4, Directness 5 means high warmth and fully direct — a friendly person who gets straight to the point. Warmth 2, Directness 3 means businesslike and measured. Separating the two prevents a model from treating warmer as permission to bury the ask under pleasantries or more direct as license to drop all warmth.

5. 5

   Tell the model who you are writing to and why

   One sentence of context: who the recipient is, how you know them, and what the email is trying to accomplish. This is a first email to a potential vendor I was referred to — I want to come across as credible and approachable, not salesy. That context changes how the model weighs the earlier inputs and how formal it goes when inputs are ambiguous.

Tone-spec blocks for eight common situations#

The five steps above work for any email. To make the method faster, here are ready-to-copy tone specs for eight situations most professionals write often. Take the row that fits, use the register, warmth, and directness values in your prompt, add the anti-example as a constraint, and then layer in your context from step 5.

How the major platforms handle tone instructions#

The five-step spec works across any AI writing tool, but each platform interprets tone instructions with its own quirks. Knowing each one saves re-prompting time.

What to do when the draft still misses the tone#

When the first draft lands wrong, resist the instinct to re-run the entire prompt. Identify which of the four building blocks drifted and correct that one thing.

• Draft too formal: the model defaulted to high-register vocabulary. Add use contractions and shorter sentences and drop the register label by one step.
• Draft too cold: warmth was not set or was too low. Raise the warmth dial explicitly and paste a sample sentence from your sent folder that has the warmth you want.
• Too much hedging: tell the model to remove phrases like just, sorry to bother you, no rush, and if that is okay — state the ask directly. Hedging almost always needs to be removed, not added.
• Anti-examples still appearing: repeat the anti-example at the end of your prompt as a check — Before finishing: confirm you have not used this phrase. This works for stubborn patterns the model keeps defaulting to.
• Still sounds like the model, not like you: you asked for a fresh email instead of an edit of your draft. Paste your draft as a starting point and ask the model to adjust the tone of your words rather than replace them.

A faster way: tone control that lives in your inbox#

The five-step spec method works, and using it produces better first drafts than going in cold. The ongoing cost is that you rebuild the spec from scratch for each email, in a chat window away from your inbox, re-teaching the model your voice every session.

We build AI Emaily — an AI-native email client that runs tone control where the email actually lives. Tone preferences sit in a user-set Personal Context brain and per-client profiles you configure once. Every draft arrives in your register by default without you specifying it. You can still adjust the warmth or directness for any individual email, and Copilot approval is on by default — you review and sign off before anything leaves your outbox, with undo and a full audit trail.

AI Emaily runs a 7-day free trial on Pro. Start at app.aiemaily.com/signup.