Draft replies that sound like you

Most people's objection to an AI drafting their email is not accuracy. It is that the result sounds like customer service. Too many words, too polite, opening with gratitude, closing with an offer to help further.

The fix is not a longer instruction about tone. It is giving it your actual sent mail and telling it to imitate rather than to write well.

Point it at how you already answer

Your sent folder is the material. You do not have to prepare anything.

Read the last fifty emails I have sent that were replies, ignoring
anything under two sentences.

Describe how I write: sentence length, how I open, how I close, whether I
use names, how direct I am, what I never do.

Be specific and unflattering. I want the pattern, not a compliment.

Read what comes back. It is usually uncomfortable and accurate, and it is the brief for everything after.

Draft, never send

When a thread needs a reply from me, draft it in Gmail. Never send.

Match how I write, per the description above. Specifically:
  open the way I open, which is usually not with thanks
  keep it to the length I would actually have written
  no sign-off flourish I do not use

If the reply needs a fact I have not given you, leave [?] in place rather
than writing around it.

Do not reply at all to anything where the answer commits me to money, a
date, or a decision I have not already made. Flag those instead.

[?] beats a graceful sentence that avoids the gap. A draft that smoothly talks around a fact it does not have is the one you send without noticing, and the missing thing surfaces a week later in the customer's reply. A visible marker costs you five seconds and cannot be skimmed past.

Read the first twenty before you trust any

You are checking one thing: would you have sent this? Not "is it good". Good and yours are different, and only the second one matters here.

When a draft is wrong, say why in a sentence and have it update the description of your voice. Ten of those corrections is worth more than any amount of prompt writing up front.

Keep the boundary even once you trust it

Drafting is safe because you press send. That property is doing most of the work here, and it is worth keeping for longer than feels necessary.

If you eventually let it send, take the narrowest slice: replies that only confirm something already agreed. Never the ones that commit you.

What good looks like

You open Gmail and half your replies are already written, in your words, and your edit is a sentence rather than a rewrite.

The failure to watch for is drift toward the pleasant middle. If drafts start sounding more polite than you are, refresh the voice description against your recent sent mail rather than adding adjectives to the prompt.