3. Getting AI to Write Something You Would Actually Send

This is the thing most people try first, and it is where most people quietly give up. You ask for an email, you get back something that opens with "I hope this message finds you well," and you think: I would never send this, I may as well write it myself.

You are right. You should never send that. But the problem is not the AI. It is that you asked for "an email" and gave it nothing to work with, so it produced the statistical average of every email ever written, which is exactly as bland as that sounds.

Here is the fix, and it is 3 things.

The 3 things a writing request needs

Every writing prompt that works has these. Every one that fails is missing at least one.

1. The facts. Bullet points are fine. Ugly is fine. It cannot invent what happened, and when you leave a gap it will fill it with something generic or something wrong.
2. The voice. Two or three adjectives. "Plain, direct, friendly, no corporate filler" does more work than a paragraph of explanation.
3. The constraints. Length, and any specific thing you do not want. This is where you kill the filler before it appears.

A real one, start to finish

A genuinely unpleasant email: telling a client you are going to miss a deadline, and it is your fault. Here is the actual prompt, copy-pasteable:

You are helping me write a work email. Keep my voice:
plain, direct, friendly, no corporate filler.

I need to tell my client we are going to miss Friday's
deadline. Here are the facts:
- The delay is our fault. Our designer was out sick 4 days.
- Realistic new date is the following Wednesday.
- They have a launch event 2 weeks after that, so this does
  not break their launch.
- I do not want to offer a discount.

Write the email. Keep it under 150 words. Only apologize once.

Notice what is in there. Four facts, including the awkward one. A voice instruction. A length. And one specific thing to avoid, because over-apologizing is the exact failure mode of an email like this.

We ran it. This came back:

Hi [Client name],

We're going to miss Friday's deadline. Our designer was out sick for four days, and we can't ship quality work without more time.

We'll have it ready for you next Wednesday instead. That still gives you two weeks before your launch event, so your timeline hasn't changed.

I know this isn't ideal. We're pushing hard to keep the delay short, and I'll send you a progress update by tomorrow so you can see where we are.

Let me know if Wednesday works for you.

That is 91 words and it is sendable. You would change the bracketed name and maybe one phrase.

Why that one worked

Go through it against what we asked for. It leads with the bad news instead of burying it in paragraph three. It gives the reason in 1 clause and does not grovel. It does the reassurance work by pointing at their launch date, which is a fact we supplied and it used correctly. It apologizes exactly once, in the phrase "I know this isn't ideal."

Then it did something we did not ask for: it offered a progress update tomorrow. That is a genuinely good instinct for this email, because the client's real worry is not Wednesday, it is whether you have this under control.

The follow-up that does most of the work

You will rarely love the first draft. Do not start over. Say what is wrong in the same conversation:

Warmer. And cut the last line, it sounds like I am
asking permission.

It keeps everything else and changes that. This is the single biggest behavior difference between people who find AI useful and people who do not. The people who find it useless write one prompt, judge the result, and close the tab. The people who find it useful treat the first output as a starting point and spend 3 rounds on it.

Four more writing jobs, with the prompt

Declining something without burning the bridge:

Write a short reply declining this freelance project.
Reason: I am booked until November and this needs to
start in 2 weeks. I want to work with them later.
Warm but not apologetic. Under 80 words.

Turning your own mess into something readable:

Here are my raw notes from a meeting. Turn them into a
summary I can send to 6 people who were not there.
Group by topic, pull out decisions and action items
separately. Keep it factual, do not add anything I
did not write.

[paste your notes]

Fixing something you already wrote:

Here is a paragraph I wrote. Keep my voice and my
points. Just make it clearer and about 30% shorter.
Tell me what you cut and why.

[paste your paragraph]

That last clause matters. Asking it to explain the cuts turns an edit into a lesson, and you start making those cuts yourself.

The awkward one you have been putting off:

I need to ask a friend to pay back money they borrowed
5 months ago. $400. They have not mentioned it. We are
still close and I do not want this to be weird.
Write 3 versions: very casual, direct, and formal.

Asking for 3 versions is a good move on anything where you do not know the right register. You pick one, or more often you take the opening of one and the ending of another.

What it is bad at, honestly

It cannot do jokes that land, because humor is specific to people who know each other and it does not know your people. It cannot do genuine emotional weight in a condolence note, and you should not ask it to. And it defaults to a mid-Atlantic business register that reads as slightly American and slightly corporate no matter what you do, so if your workplace writes in a distinct way you will need to show it examples. That is lesson 6.

When it still sounds like a robot

You will hit this. The facts are all there, the length is right, and it still reads like a brochure. Three fixes in the order we try them:

Read that back and cut every phrase that a human
would not say out loud. Then show me what is left.

This one is oddly effective. "Please do not hesitate to reach out" does not survive being said out loud, and neither does most of the filler.

Rewrite it the way I would say it to them in person,
then tidy it up just enough to be an email.

Spoken register is looser and warmer, and starting from there and tightening gives a better result than starting from written register and trying to loosen it.

Here are 2 emails I actually sent. Match this voice.
[paste]

If the first 2 do not work, this always does. It is lesson 6 and it is the real answer to the robot problem.

Six mistakes worth skipping past

1. Asking for "professional". It already thinks it is being professional. That is the problem. Ask for plain, or direct, or warm.
2. Not saying who it is to. The same facts go to a client, a colleague and your boss in 3 different registers.
3. Regenerating instead of correcting. Rerolling gives you a different average. Saying "warmer, cut the last line" gives you your email.
4. Leaving in the brackets. Everyone has sent one with [Client name] still in it. Search for the square bracket before you send. Every time.
5. Letting it soften bad news into vagueness. Its instinct is to cushion. If the message is "we are late", check that a reader skimming it would actually know that.
6. Using it for the message that needed to be a phone call. It will happily write you a beautifully worded email that should not have been an email.

Do this one now

Pick an email or message you have been avoiding. Genuinely avoiding, the one sitting in your head. Write the facts as bullet points, add a voice line and a word count, and see what comes back.

The task you are avoiding is usually the one where this helps most, because the reason you are avoiding it is that you do not know how to open it.

Next

Lesson 4 is the one that changes what this tool is for. So far you have been asking it to produce things. Next you paste in something you already have, a contract, a bill, a report, and start asking it questions about your own documents.