The prompt
Copy it whole, replace every bracketed variable, then paste it as your first message.
You are my drafting partner for one email. I direct, you propose, I judge, I own what gets sent. Read everything before you reply. ## 1. The situation - Recipient and relationship: [RECIPIENT AND RELATIONSHIP] - The ask, with its deadline: [THE ASK] - The facts: [THE FACTS] - History you cannot see: [HISTORY THE MODEL CANNOT SEE] - Must not say: [MUST NOT SAY] - My usual register with this person: [YOUR USUAL REGISTER] - Length ceiling: [LENGTH] ## 2. Rules for the email - The ask appears by the end of the second sentence. The reader must know what I need before they know why. - State each fact once. Do not restate facts as reassurance. - Apologize at most once, only if something is actually my fault, and in one sentence without "sincerely", "deeply", or "any inconvenience". If nothing is my fault, do not apologize; acknowledge instead. - No blame, including implied blame through passive voice. State what has and has not happened as plain fact. - The deadline is on its own line or in its own short sentence, with the actual day. - Offer at most two options if a decision is needed. Say which I recommend, in one sentence, if I have a view. - End with the next step, not with a courtesy. No "please let me know if you have any questions", no "thanks in advance", no "I hope this finds you well" at the start. - Subject line: under 60 characters, states the ask or the decision, no "Quick question", no "Update". - Match the register in section 1. Do not add warmth the history does not support, and do not remove warmth it does. - Do not invent facts, names, dates, or figures. Where the email needs something I have not given, write [VERIFY: what is needed]. Banned anywhere: "I hope this email finds you well", "Just checking in", "circling back", "touch base", "as per", "at your earliest convenience", "going forward", "I wanted to reach out", exclamation marks, more than one "sorry". ## 3. How we work Before drafting, ask me at most one question, only if its answer would change the email. Then give me the draft: subject line first, then the email, then stop. No commentary. When I say "change only X": return only the changed sentences marked OLD and NEW. When I say "options for the opening" or "options for the ask": three versions of that part only, labeled by tradeoff in six words or fewer. When I say "shorter": cut whole sentences, never compress; list what you cut. When I say "tone check": do not rewrite. In three lines, tell me how the recipient is likely to read the email given the history, what they might resent, and what they might miss. ## 4. Begin If you have a question, ask it. Otherwise, draft.
- Best for
- Emails with a real ask: a decision, an approval, a date, a fix
- Delivering bad news without hiding it or groveling
- Messages to someone with history: a client, a manager, a vendor, a colleague
- Writers who know what they want to say and keep softening it into mush
- Not for
- Mass emails and marketing sends. Different rules. Use the Professional Document prompt or your email tool's guidance
- Anything containing information you are not permitted to paste into a third-party tool. Describe it in general terms or write by hand
- Emotional messages to people you love. Write those yourself
Email is the writing most people do most, and the writing AI assistance most reliably makes worse. The drafts are polite, long, and shapeless. They apologize three times. They bury the request under context. They end with 'Please let me know if you have any questions', which is a sentence that has never once been needed.
The problem is that an email is mostly relationship, and the model cannot see the relationship. It does not know this client has already had one slip, that your manager hates preamble, that the last three messages in the thread went unanswered. This prompt asks for that history up front and treats it as a constraint. Then it enforces the shape of an email someone will read on a phone: the ask early, the facts once, the deadline unmissable, the tone matched to the history rather than to a default of nervous courtesy.
Use it for the emails that matter. For the rest, the Bounded Edit prompt on your own draft is faster.
Variables
What to put in each bracket. Required ones are marked. Thin answers here are the usual cause of generic output.
[RECIPIENT AND RELATIONSHIP]RequiredWho they are to you, and the state of the relationship right now.
ExampleA client of two years, project sponsor, generally warm, currently stretched thin and slow to reply
[THE ASK]RequiredWhat you need from them, in one sentence, with the date you need it by.
ExampleA decision by Friday on whether to delay launch two weeks or cut the reporting feature
[THE FACTS]RequiredWhat happened and what is true, briefly. Only what the email needs.
ExampleTheir data team has not delivered the export spec; without it we cannot build reporting; two-week slip or cut the feature; everything else is on track
[HISTORY THE MODEL CANNOT SEE]RequiredPrior slips, promises, tensions, what has already been said, what tone has worked before.
ExampleWe slipped once before in March, our fault. I said then it would not happen again. This one is not our fault, but I do not want to sound like I am pointing at them
[MUST NOT SAY]Things that must not appear, for legal, political, or relationship reasons.
ExampleDo not blame their data team by name. Do not promise a date for the reporting feature. Do not mention the March slip
[YOUR USUAL REGISTER]RequiredHow you normally write to this person, with a short example if you have one.
ExampleFirst names, short paragraphs, no exclamation marks, I sign off with just my name
[LENGTH]A ceiling in words. Most emails with one ask should be under 150.
ExampleUnder 150 words
Why it is built this way
Each block does one job. Knowing the job lets you shorten or adapt the prompt without breaking it.
History the model cannot see
This is the variable that separates a usable draft from a generic one. Prior slips, past promises, and the current temperature of the relationship change every sentence, and no model can infer them.
Must not say
Emails get forwarded. An explicit list of what cannot appear keeps the draft safe when the model's instinct to be thorough would include the thing you are managing around.
The ask by sentence two
The single rule that fixes most AI-drafted email. Context-first structure is the model's default and the reader's enemy.
One apology, or none
Models apologize as a courtesy reflex. Limiting it to fault, once, in one plain sentence, keeps you from sounding either guilty or insincere.
No passive-voice blame
'The spec has not been delivered' blames quietly. 'We do not yet have the spec' states the fact. The prompt names the trick so the model avoids it.
The banned phrases list
Every item is a sentence readers skip or wince at. The list is short because it only needs to cover the ones models reach for most.
Tone check as a separate move
A read from the recipient's side is the most useful thing the assistant can offer on email, and it should never come bundled with a rewrite you did not ask for.
Example run
A representative first response, with the input summarized. Your results will differ by model and material.
Filled for the delayed launch scenario: a two-year client, a Friday decision between a two-week slip and cutting the reporting feature, a missing export spec on their side, a prior slip in March that must not be mentioned, and a plain first-name register.
Subject: Launch decision needed by Friday Hi Dana, I need a decision from you by Friday, 24 October: either we move launch two weeks to 10 November, or we launch on 27 October without the reporting feature and add it afterward. The reporting build depends on the export spec, which we do not yet have. Everything else is on track and tested. My recommendation is the two-week move. Launching without reporting means your finance team works from the old process for at least a month, and I would rather not put that on them. Could you reply with your choice by end of day Friday? I will send the revised plan the same day. Sam
The ask is in the first sentence, the fact about the spec is stated once without a subject that blames, the March slip is absent, and the recommendation is one sentence with a reason the client will recognize as hers. The writer changed 'put that on them' to 'ask that of them' and sent it. The reply came Thursday.
Follow-up prompts
The second turn is where vibe writing happens. These are the short messages that keep the session moving.
Before sending anything that matters
Tone check.
When the opening is still a warm-up
Options for the opening. The ask must be in sentence one or two. Three versions, labeled by tradeoff.
When it is over the length ceiling
Shorter. Under [NUMBER] words.
When the email needs to be firmer without being cold
Change only the final paragraph. I need a clear deadline and a consequence if I do not hear back, stated as what I will do, not as a threat. OLD and NEW.
For a follow-up when they have not replied
They did not reply by the deadline. Draft a follow-up under 60 words: reference my previous email by date, restate the ask in one sentence, give a new deadline, and say what I will do if I hear nothing. No apology for following up.
When it goes wrong
Symptoms you will actually see, the usual cause, and the correction that works.
The draft is polite, long, and the ask is in paragraph three.
- Cause
- The ask variable was vague ('discuss next steps') or the facts were long enough to pull the structure toward context-first.
- Fix
- Rewrite the ask as a sentence containing a verb and a date. Shorten the facts to what the email needs. Re-draft.
It apologizes twice and uses 'inconvenience'.
- Cause
- The history mentioned fault, and the model's courtesy reflex multiplied it.
- Fix
- Reply: 'One acknowledgment, one sentence, no apology words. Change only that.' Check that the one remaining sentence is for something that is actually yours.
A date, name, or figure appears that you did not give.
- Cause
- The model needed a specific to make the email sound decisive.
- Fix
- Replace with the real value or a [VERIFY]. Never send an AI-drafted email without checking every date and number against your own records.
The tone is right but it sounds like a template, not like you.
- Cause
- The register variable was a description rather than an example.
- Fix
- Paste a short email you actually sent to this person and say 'Mine. Match this.' It takes ten seconds and fixes most of it.
Before you publish
The prompt cannot do these. You can.
- Every date, name, figure, and attachment reference is correct. Check against the thread, not against memory.
- Nothing in the 'must not say' list appears, including by implication.
- Read it as the recipient, in their current mood, on their phone. If the ask is not clear in five seconds, move it up.
- Confirm you are comfortable with it being forwarded to the recipient's boss. Emails travel.
- If your workplace has a policy on AI-assisted communication, follow it. Most do not require disclosure for drafting help, but some client contracts do.
Questions
Is it safe to paste work emails into an AI assistant?
Only if your organization permits it and the content is not confidential, regulated, or covered by a client agreement. When in doubt, describe the situation in general terms rather than pasting the thread, or use a tool your organization has approved. The prompt works with summaries.
The prompt is longer than most of my emails. Is it worth it?
For a routine email, no. Write it yourself and use the Bounded Edit prompt if you want a second look. For an email where the relationship or the outcome matters, the five minutes spent on the history and the must-not-say list is where the value is.
How do I make an email firmer without sounding aggressive?
State facts without adjectives, give a specific deadline, and say what you will do if you do not hear back, framed as your action rather than their consequence. The firmer-ending follow-up above produces exactly that. Firmness is specificity, not tone.
Can it write the whole thread, including replies?
Each email is its own session with its own history. For a reply, add the message you are replying to as part of the facts and update the history variable with what has changed. Do not ask it to predict the other side's replies. That is where invented facts come from.