AI writes the first draft, you write the final one

The hardest part of an unpleasant email isn't the content — it's the first sentence. You know exactly what you want to say — decline, chase, apologize — and yet you sit in front of a blank window for twenty minutes hunting for an opening that won't sound like an excuse or an attack. That delay is expensive: the email stays half-written, goes out late, and in the meantime haunts you all day.

The fix isn't letting AI write your emails for you. It's moving from a blank page to a draft in progress: generate something with the right structure and an acceptable tone, then rewrite it in your own voice. The whole difference between “write an email to the client” and a usable result lives in the request — and that's the core of this guide.

We'll go through the four ingredients of a good brief, ready-to-use requests for the four most unpleasant email types, working with a full thread (where most mistakes hide), and how to hand the model your own style so the draft doesn't read like it came out of a brochure. At the end there's a check before you send and rules for sensitive messages. One principle sits above all of it and will keep coming back through this guide: AI proposes, a human always sends.

A typical scenario

Jane runs projects at an agency and gets thirty to forty emails a day. Most of it is routine, but three to five are the ones where she gets stuck: a client wants a deadline that can't be met, a vendor has gone quiet for the third week straight, a colleague made a mistake that needs escalating, and every so often she has to own up to something her team got wrong.

It used to cost her twenty minutes and a half hour of dread per email. She'd open the window, type “Hi, thanks for your message…”, delete it, try again, go get coffee, come back, finish a paragraph, read it, and realize it sounded irritated. She'd send it that evening or the next day. Over a week, that added up to two to three hours and a nagging feeling of something left unfinished.

Now she has a saved prompt for each of those four types. She pastes in the thread, fills in three lines of brief (what she wants, what tone, how long), and has a draft in fifteen seconds. Ten minutes of rewriting it in her own voice — because the model can't and shouldn't decide how much ground to give and what to promise — and the email goes out within half an hour of her reading it. The saved time is one thing; the bigger win is that unpleasant emails stopped being a pile she had to work up the nerve to face.

The rest of this guide describes exactly what's in those saved prompts.

Phase 1: the brief method, four things in three lines

Why “reply to this email” isn't enough

When the model gets only the incoming message, it has to guess what you actually want. And it guesses predictably wrong: it's maximally accommodating, agrees, offers concessions you never intended to give, and writes something twice as long as it should be. What comes out are emails that box you into a position you never chose.

A good request holds four things. Remember it as who — what I want — tone — length.

Who is the relationship context. It's not a name — it's a type: a longtime client, a new client in their first week, a vendor you're still building trust with, a colleague two levels up, an opposing party in a dispute. The model infers formality, how much needs explaining, and what it can safely leave unsaid from that.

What I want is the outcome, not the topic. Not “write about the deadline,” but “I want them to accept the shift to October 15th without treating it as a failure.” A goal phrased as the recipient's outcome is the single strongest line in the whole brief.

Tone is one or two precise words, not adjectives picked at random. “Matter-of-fact, no apologies,” “warm but not backing down,” “friendly, informal.” Avoid “professional” — it means nothing.

Length is the most underrated part. Without it, the model writes long. Give a sentence or paragraph count: “max 6 sentences,” “three short paragraphs,” “one paragraph and one closing question.”

The basic brief template

This prompt covers eighty percent of everyday emails. Save it first.

Write a draft reply to this email. I'll still be editing the
reply, so I want a draft, not a final version.

INCOMING MESSAGE:
[paste the email text]

WHO I'M REPLYING TO: [relationship — e.g. a client I've
worked with for three years; formal tone]
WHAT I WANT TO HAPPEN: [the outcome for the recipient, not
the topic]
FACTS THAT MUST BE INCLUDED: [dates, numbers, terms — only
what's actually true]
WHAT MUST NOT BE INCLUDED: [what not to promise, not to
admit, not to touch]
TONE: [e.g. matter-of-fact and warm, no apologetic openers]
LENGTH: [max 6 sentences]

Rules:
- don't make up dates, prices, or circumstances; if something
  is missing, write TO ADD: [what] at the end
- don't open with phrases like “thank you for your email” or
  “hope you're doing well”
- no closing offers like “please don't hesitate to reach out
  with any questions” unless I explicitly ask for one
- suggest a subject line too, if it should change

It returns a draft you can send after two edits. The most important line is “what must not be included” — without it, models like to tack on an extra accommodating sentence (“of course we can make that work”) that commits you to something you never meant to promise. And take the rule about making things up seriously: the model will fill in a date or an order number with total confidence, even if it's never seen either.

Three versions instead of one

For sensitive messages, it's worth generating three positions at once and choosing between them. It saves a round of rewriting, and more importantly, it shows you where your own line actually sits.

For this situation, write three versions of a reply, not one:

SITUATION: [description]
INCOMING MESSAGE: [paste it]
WHAT I WANT: [outcome]

A) SOFT — maximally accommodating, protects the relationship,
   gives ground on [what's negotiable]
B) NEUTRAL — matter-of-fact, no emotion, just facts and the
   next step
C) FIRM — clearly holds my position, polite, but leaves no
   room for negotiation

For each version, add one sentence on what I risk by sending
it, and how the recipient is likely to react.
All three max 8 sentences. Don't mark any of them as
recommended.

It returns three drafts and a reaction estimate. That last part is the main reason to use this prompt — you'll often find that the version you were leaning toward triggers five more back-and-forth emails. Treat the reaction estimates as food for thought, not a forecast — the model doesn't know the recipient.

Phase 2: four hard emails

These four types make up most of the messages that sit half-written and stuck. Each has its own logic and its own trap.

Rejection

The trap in a rejection is padding: the longer the preamble, the worse it lands. A good rejection is short, says no in the first two sentences, explains why in one sentence, and offers an alternative if one exists. Don't explain yourself three times over — the reader reads uncertainty into that and starts negotiating.

Write a rejection of this request.

REQUEST: [paste the message]
WHO: [relationship]
REAL REASON FOR THE REJECTION: [the actual reason]
WHAT I CAN SAY OUT LOUD ABOUT THE REASON: [what's shareable —
leave the rest unwritten]
ALTERNATIVE I'M OFFERING: [or “none”]

Structure:
1. One sentence acknowledging I understand the request — no
   flattery.
2. A clear no in the second or third sentence. Not “that will
   unfortunately be complicated,” an actual rejection.
3. One sentence with the reason. Not one extra apology.
4. An alternative, if one exists, as a concrete offer with a
   date.

Max 6 sentences. Don't use the conditional where it's actually
a decision (“we wouldn't be able to” → “we won't”).
Don't open with an apology.

It returns a rejection that can be sent unedited more often than the other types. Check point 3 — models tend to smear the reason across three sentences and add “I understand this isn't the news you were hoping for.” That sentence sounds nice and undercuts the whole email.

Follow-up and reminder

This is where people mess up on both ends: either the message is so soft nobody notices it, or so sharp that an ordinary delay turns into a conflict. A working follow-up doesn't assume bad intent, states concrete facts (what was sent when, what was agreed), and ends with one question or one deadline.

Write a follow-up. They haven't replied and I need to move
this forward.

WHAT HAPPENED: [what I sent, when, what was agreed]
WHICH FOLLOW-UP IS THIS: [first / second / last before
escalating]
IMPACT OF THE DELAY: [what specifically gets held up or
costs more]
WHAT I NEED: [a reply / delivery / a signature] BY [date]
RELATIONSHIP: [relationship and how much I can push]

Rules:
- don't assume or imply bad intent
- state facts and dates exactly as I gave them above, don't
  do any math
- no passive-aggressive phrasing (“as I mentioned before,”
  “once again,” “still waiting”)
- end with one specific request with a date, not an
  open-ended ask
- for the second follow-up and beyond, add a sentence on what
  happens if no reply arrives — matter-of-fact, not a threat

Max 5 sentences. Suggest a subject line that makes the
point clear without opening the email.

It returns a follow-up that's still polite but impossible to ignore. The ban on phrases like “as I mentioned before” is deliberate — those are exactly the sentences a recipient reads as a rebuke, and they harden the conversation faster than you'd want. When you're sending a lot of these, it's worth setting up a system: a follow-up watchdog tracks sent emails with no reply and drafts the follow-up itself.

Escalation

Escalation is the hardest type, because more people read it than are on the To line, and it stays in the archive. There's one rule: describe the problem and the impact, not the people. The moment an email evaluates a colleague, it stops being about the issue and starts being about you.

Write an escalation email. I need to move this up because it
hasn't budged at my level.

PROBLEM: [what specifically isn't working — facts, not
impressions]
WHAT I'VE ALREADY TRIED: [steps and dates]
IMPACT: [on the deadline, the budget, the client, the risk]
WHO I'M WRITING TO: [role, how far removed from the problem]
WHAT I NEED FROM THEM: [a decision / resources / authority —
specifically]
BY WHEN: [date and why that date]

Rules:
- no evaluation of individuals, no names in the role of
  culprit; describe processes and facts
- the “what I've tried” section has to be verifiable, don't
  pad it
- offer 2 solution options and note what each one requires
- tone: calm, matter-of-fact, no urgency and no drama
- structure: 3 short paragraphs (status / impact / what I
  need) plus bullets with the options

Attach a subject line variant that makes clear this needs a
decision, not just information.

It returns a structured email that still reads well three months later. The most valuable part is asking for two solution options: an escalation that only brings a problem reads as a complaint; one with options reads as work. Before sending, check the “what I've tried” section — it needs to be accurate down to the day, because that's exactly the paragraph someone will verify.

Apology

The opposite rule applies to an apology compared to a rejection: the sooner the admission, the better. The structure has four parts — what happened, what it means for the recipient, what we're doing about it (with a date), and what we're doing so it doesn't happen again. Missing the third and fourth parts is why most apologies don't work.

Write an apology email. This happened on our end.

WHAT HAPPENED: [factually, without explaining the
circumstances]
IMPACT ON THE RECIPIENT: [what it cost them — time, money,
extra work]
WHAT WE'RE ALREADY DOING: [concrete steps and who's doing
them]
BY WHEN IT WILL BE FIXED: [date]
FUTURE SAFEGUARD: [what we're changing so this doesn't
happen again]
WHAT I'M NOT OFFERING: [discounts, compensation, or promises
I don't have approval for]

Rules:
- apologize in the first or second sentence, clearly and
  once; don't repeat it
- no “if we caused any inconvenience” — either we did, or we
  didn't
- don't blame circumstances, vendors, or the recipient
- don't offer any compensation beyond the “what I'm not
  offering” section
- end with one concrete next step, not a plea for
  understanding

Max 8 sentences, tone serious and calm, no self-flagellation.

It returns an apology that owns the mistake without collapsing into it. The “what I'm not offering” section exists for a practical reason: models have a strong tendency to offer compensation or a discount because it sounds accommodating — and that commits you to something you never had approval to promise. Read the draft out loud once more; with apologies, a wrong tone is easier to catch by ear than by eye.

Phase 3: working with the full thread

Why the last message isn't enough

The most common source of awkward replies is feeding the model only the last email. In a long thread, half of what matters happened earlier: what you already turned down once, which deadline moved, what a colleague on cc promised, what exception applies. A model that only sees the last message has no way to know any of that — and writes a reply that contradicts what you wrote a month ago.

Paste in the whole thread, even if it's long. And add an instruction that bans the model from filling gaps on its own:

Below is the full email thread. Reply on my behalf to the
last message.

THREAD (oldest at the top):
[paste the whole thread, including From/To/Date headers]

I AM: [your name and role in the thread]
WHAT I WANT FROM THIS REPLY: [outcome]
TONE: [tone]
LENGTH: [sentence count]

Rules I insist on:
- stick to the facts in the thread; don't infer, don't
  generalize
- don't promise anything I haven't already promised in the
  thread
- if messages in the thread contradict each other, flag the
  contradiction and ask what applies — don't decide it for me
- respect what I've already declined; don't reopen it
- take figures (dates, amounts, numbers) verbatim from the
  thread, don't recalculate them

Above the reply, write 3 bullets: what's still open in the
thread, what the recipient wants from me, and what I want
from them.

It returns a reply plus three bullets worth reading on their own — they're often the fastest summary of a long thread you'll get anywhere. Watch the rule about promises in particular — it's the most common spot where the model adds an accommodating sentence nobody approved.

Extracting the thread before replying

For long, contested threads, it's worth splitting the work into two steps: first have the facts of the thread listed out, only then write the reply. That keeps an assumption from slipping into the email.

Don't write a reply yet. Pull the facts out of this thread
for me:

[paste the thread]

1. COMMITMENTS — who promised what to whom, with a date and
   the exact sentence it comes from.
2. DEADLINES — every date mentioned in the thread and what's
   tied to it. Flag which have already passed.
3. NUMBERS — amounts, counts, ranges, including who stated
   them and when.
4. OPEN QUESTIONS — what nobody has answered.
5. CONTRADICTIONS — places where two messages disagree.

Don't calculate or round anything. Where the author of a
claim isn't clear, write UNCLEAR. Don't draw any conclusions.

It returns an overview that the reply practically writes itself from. Point 1 with citations is the part that makes the extra minute worth it — on a contested thread, you need to know what you're bound by before you start phrasing anything. Spot-check the quotes against the original, though — the model can unintentionally strengthen a sentence.

Phase 4: making it sound like you

Pull a style profile out of your own emails

A draft that's factually correct and reads like a corporate flyer still gets rewritten in full. It's faster to extract how you actually write once, and add that description to your requests — or save it into a project.

I'm attaching [8] emails I wrote and sent myself. They're
ordinary work messages, nothing especially polished.

[paste the emails]

Describe my style so it can be used to write an email that
sounds like me:

1. OPENINGS: exactly how I open (quote 3 examples)
2. CLOSINGS: how I close and sign off (quote 3 examples)
3. FORM OF ADDRESS: formal or informal, first names, titles
4. SENTENCES: length, whether I write in bullets or flowing
   prose
5. TYPICAL WORDS: phrases that recur
6. WHAT I NEVER USE: phrasing I clearly avoid
7. DEGREE OF POLITENESS: where I'm terse and where I soften

Work only from the attached emails. Where the sample's too
small for a conclusion, say so instead of stating it as fact.
The output needs to be specific enough to use as an
instruction.

It returns a description you can paste into every email prompt from here on — or better yet, into persistent context so it attaches on its own; how to do that is in the tip on projects as persistent memory. Pick emails from ordinary day-to-day traffic for the sample, not the three you polished carefully: you want your actual voice, not your Sunday-best one.

A sample works better than a description

Even without a style profile, a simple shortcut applies: attach one email of yours of a similar type and say “like this.” For format and tone, one concrete sample beats five sentences of description. If you regularly handle a certain type of message, keep one sample email for it saved and attach it every time.

Saved prompts instead of writing the request again

Once a prompt has worked twice, save it — in notes, in text expansion snippets, in a prompt library. The request then shrinks down to filling in three brackets. Classic Gmail templates work on the same logic: AI suits emails where the content differs every time, a template suits the ones where it's always the same.

I used this prompt and the result was good:

[paste the prompt]

Turn it into a general-purpose template I'll reuse for the
same type of email:
- replace the variable parts with bracketed placeholders that
  say what goes there (not [text], but [reason for the
  rejection, 1 sentence])
- keep the rules that worked verbatim
- add a line for tone and length so those can vary
- finish with 4 points I should check in the draft before
  sending

The template has to work for a situation you haven't seen.

It returns a prompt ready to save. Don't skip that final checklist — after a few weeks of use, you'll stop reading drafts closely, and that's exactly when a sentence you didn't want to send slips through.

Phase 5: what to check before you hit send

The final pass

However good the draft is, it goes through one more read. This prompt does a check your own eyes won't catch on a third read-through — but never let it rewrite the draft.

Here's a draft email I'm about to send:

[paste the draft]

CONTEXT: [who it's going to, what the goal is]

Don't rewrite it. Just list findings in five categories:

1. COMMITMENTS — sentences that commit me to something
   (a deadline, a price, scope, a guarantee). Quote them
   verbatim.
2. TONE — spots that might land sharper or softer than I
   probably intend. Quote and explain why.
3. UNSUPPORTED FACTS — numbers, dates, and facts I haven't
   backed up anywhere in the request.
4. SENSITIVE DATA — personal data, internal information, or
   third-party details that might not belong here.
5. AMBIGUITIES — sentences that can be read two ways.

At the end, write in one sentence how the message will
probably land with a recipient who's having a bad day.

It returns a list where point 1 and the closing sentence are the most useful. Commitments quoted verbatim are the fastest way to catch a draft promising more than you intended. And reading it “through the eyes of a recipient having a bad day” surfaces exactly the phrasing that causes unnecessary conflict.

The final pass is always yours

Rewriting the draft isn't a delay — it's the actual work. The proven approach: read the draft, verify the facts, then rewrite the first and last sentences in your own words — those two carry most of the tone — and go through the rest. Anything you'd never say yourself, cut it, even if it sounds professional.

And a rule with no exceptions: a human always sends. Even in automated setups that prepare drafts in advance (see a morning routine with reply drafts), nothing goes out without a human reading it first. The difference between a draft and a sent message is your entire reputation.

Sensitive emails and other people's data

Email inboxes are full of things that don't belong in an open chat: personal data, internal figures, contract terms, health information, HR matters. Three rules apply.

Internal and sensitive messages only in a paid account with contractual data protection, never in a free tier, and never in a tool where you don't know who has access to the data. In a company setting, ask whoever manages this which tool is approved — and if none is, don't put a sensitive email into AI at all.

Anonymize anything you don't need for phrasing. Third-party names, phone numbers, contract and account numbers, addresses — the model doesn't need any of it to write the reply.

I'm attaching an email I need help replying to. Before that,
make a working version of it:

[paste the email]

- replace people's names with their role ([client], [colleague
  from production])
- replace company names with [other party], [vendor]
- remove phone numbers, addresses, email contacts, contract
  numbers, payment reference numbers, and bank details
- keep amounts only where they're necessary for the reply
- don't otherwise change the content, tone, or structure

Return the cleaned version and a list of what you replaced
below it.

It returns a version you can safely keep working with. Check footers and signatures — that's where contacts tend to hide and get overlooked. And the third rule: someone else's confidential message isn't yours just because it landed in your inbox. Before you paste it anywhere, consider whether the sender would expect that.

Common mistakes

• Pasting in only the last message from a thread. The model has no way to know what you already declined or promised three weeks ago — and writes a reply that contradicts it. Paste in the whole thread, even if it's long.
• Forgetting to specify length. Without a limit, you get an email twice as long as the situation calls for. For unpleasant messages, length signals uncertainty: the longer the explaining, the weaker the position looks.
• Leaving a commitment in the draft you never meant to make. Models are accommodating and love to add a sentence about how it's sure to work out. List out every commitment in the draft, verbatim, before you send it.
• Trusting the numbers and dates in the draft. The model fills in deadlines, amounts, and order numbers with complete confidence. Every figure needs to come from the thread or your own request, never from the draft itself.
• Sending without rewriting it in your own voice. Recipients spot a generic email faster than you'd think, and on sensitive messages it reads as distant exactly where a human touch matters.
• Putting internal and personal data into an open chat. Sensitive emails belong only in a paid account with contractual data protection, and even there, anonymized — nobody needs a third party's name and phone number to write the reply.
• Automating the sending. Preparing a draft in advance is safe; sending without reading it isn't. A human always sends, no exceptions.

The best tools

• Claude or ChatGPT in a browser — paste in the thread and the brief, get a draft; the most control over the request and the best place for saved prompts.
• An AI assistant built into your email client (Copilot in Outlook, Gemini in Gmail) — the draft appears right in the compose window, no copy-pasting; convenient for quick replies, weaker where you need a detailed brief.
• A project with your style and templates — style profile, sample emails, and saved prompts in one place, so the context attaches itself.
• Text expansion snippets — to unfold a saved prompt with two keystrokes; the fastest way to turn a good request into a habit.
• Templates in your email client — for emails that are identical every time; save AI for the ones where the content changes each time.
• Inbox triage — so drafts only get written for what actually needs your reply.

What you get out of it

• Time: for unpleasant emails, the twenty minutes spent hunting for the first sentence shrinks to fifteen seconds of generating plus ten minutes of editing. At three such messages a day, that's roughly half an hour a day — about two days per quarter.
• Money: faster replies on proposals and follow-ups shorten how long a deal or an invoice sits open; a reminder sent on the due date works noticeably better than one sent two weeks late.
• Peace of mind: the pile of postponed “hard emails” disappears. Most of the mental load isn't the writing — it's the putting off.
• Quality: a consistent tone across your communication and fewer sentences that commit you to something you didn't mean to say. The pre-send check catches exactly what a third read-through misses.

Pro tip

Next time you have to write an unpleasant email, write a raw version first, in your own words — regardless of tone, angry if that's how you feel — and only then hand it to the model with the instruction: “this is my unedited draft; keep every fact and my position intact, only change the tone to matter-of-fact and polite, and trim it to six sentences; don't add anything, don't promise anything extra.” You'll get an email that's still unmistakably yours, just without the sharp edges, and you skip both the blank page and someone else's phrasing.

And the closing rule: a draft is raw material, not a message. You verify the facts, you approve the tone, and you click send. The model can save you the start and the structure — it can't take on the responsibility for what goes out under your name.