KeyMap
Prompt engineering course
Lesson 03beginner7 min read

Email, Slack & hard conversations

Prompts for the writing you do every day — replies, follow-ups, saying no, chasing people, and the message you are dreading.

EmailSlackDaily use
Give the model the thread and your rough intent in blunt words. Polishing your blunt intent is the job; inventing your intent is not.

The pattern that covers most email

Nearly every useful email prompt is the same shape: here is the thread, here is what I actually want to say in unvarnished words, make it land well.

The unvarnished part is the trick. Write your real feeling — “this deadline is not happening and I am annoyed they moved it twice” — and let the model handle diplomacy. If you pre-soften your intent, you get a reply that is vague about a thing you needed to be clear about.

Reply to a thread
Draft my reply to the thread below.

What I actually want to say, bluntly: [your real, unfiltered intent]

Relationship: [e.g. external client, six months in, generally good]
Constraints: keep it under 120 words, no bullet points, warm but not apologetic. Do not thank them for their patience.

Thread:
---
[paste]
---
The blunt-intent line carries the meaning; everything else controls delivery. The specific ban on one cliché works better than a general 'sound human'.
Follow-up that isn't annoying
I sent the email below 8 days ago and got no reply. Write a follow-up that gives them an easy way to respond without embarrassment, restates the one thing I need, and includes a specific date. Under 60 words. No 'just bumping this' and no 'circling back'.

Original email:
[paste]
Naming the failure modes you have seen ('just bumping this') is the one time negative instructions earn their place.

Saying no, pushing back, delivering bad news

This is where prompting pays for itself, because the hard part is not the words — it is being clear and kind at the same time while you are stressed.

The key input is the thing you are protecting. A no that explains what you are protecting reads as a priority; a no that explains nothing reads as a brush-off. Tell the model what it is.

Decline without burning the bridge
Help me say no to the request below.

Why I am saying no, honestly: [real reason]
What I am protecting: [the commitment this would displace]
What I can offer instead, if anything: [alternative, or 'nothing']
Relationship I want to keep: [context]

Write it short, decisive, and without three paragraphs of apology. One clear no, one honest reason, one alternative if I have offered one.

Request:
[paste]
'One clear no, one honest reason, one alternative' is a structural constraint. It stops the model padding a decline into a hedge.
Pressure-test before you send
Below is a message I am about to send. Do not rewrite it. Tell me:

1. How the recipient will most likely read the tone.
2. The single sentence most likely to be misread, and how.
3. What I have left ambiguous that they will have to guess at.
4. If I am wrong or unfair anywhere, say so plainly.

Message:
[paste]
'Do not rewrite it' is load-bearing — otherwise you get a rewrite instead of a review. This one is worth running on anything sent while angry.

Inbox triage

Reading is most of email. A model that reads a thread you were CC'd into halfway through will tell you what you missed faster than scrolling will.

Catch me up on a thread
I was added to this thread late. In under 100 words: what is being decided, who disagrees with whom, and what — if anything — is being asked of me. If nothing is being asked of me, say so in the first line.

Thread:
[paste]
The 'say so in the first line' escape hatch is the whole point. Most threads you get CC'd into need nothing from you, and you want that answer immediately.

A note on voice

Drafts that sound nothing like you are worse than no draft, because you spend longer un-writing them than you would have spent writing. The fix is a saved voice sample: two or three of your own emails you liked, pasted in with “match this register.”

Keep them in a note. Paste them into the tools that support persistent instructions once, and you stop needing to paste them at all.