Your edits are instructions.

The most common thing people do with an AI-written email is fix it. Cut the opening to one line. Delete the sentence that sounds like a brochure. Swap the sign-off for the one they actually use. Send. Tomorrow the brochure sentence is back, and the fix happens again.

Each of those edits felt like work. Each one was an instruction. It just had nowhere to go.

Every edit is an instruction. Most software throws it away.

The usual answer is to ask people to write their rules down. A style guide, a page of dos and don'ts, a long prompt. Some people do it. Most don't, and the ones who do tend to write the rules they believe they follow, not the ones they apply at four on a Thursday afternoon. The real rules are in the edits. Nobody wrote them down because nobody had to.

So the more useful move is to watch the edits. When the same change shows up in three separate email drafts inside a month - the opening cut to a single line, the same sentence deleted, the same sign-off added, the whole email cut by a third - the platform notices and offers to make it a rule. The offer shows its evidence: the drafts where you made that change. It does not guess at your taste. It counts.

Two edits propose nothing. Two can be a mood. Three is a habit, and a habit is an instruction you have already given, three times.

Counting sounds less impressive than understanding, and that is its strength. A system that guesses at your taste is wrong in ways that are hard to argue with, because it cannot show its working. A system that counts can show the three drafts, and you can look at them and decide whether it read you right. One of those you can trust after a week. The other you are still auditing after a year.

Two edits are a mood. Three are a habit. A habit is an instruction you already gave.

Then it waits. The rule is saved only when the person whose edits it came from says save, in their own words if they want to reword it. Nobody else can save it on their behalf. Say no and nothing is written, and the same suggestion does not come back for a month. There is only ever one suggestion open per person at a time, so it never turns into another inbox to clear. Later, ask which rules it is drafting from, and it tells you.

The limits are part of the design, not the small print. This works on email drafts written in the platform. It reads edits only from people who agreed to let their writing train how their drafts sound. Without that agreement there is nothing to read, and nothing is proposed.

That consent is not a formality either. Edits are personal. They show how someone talks, what they find embarrassing, which phrases they would never put their name to. Reading them is a privilege the writer grants, and the system behaves as if it knows that.

Good editors have always done this. They notice the third time they made the same correction, and they turn it into the house style. Good managers do it too, when they have the time to read every draft a team sends, which is never.

A correction made once is a fix. A correction noticed is a standard.

What changes is the direction of the effort. Drafts drift toward how this person writes, one rule at a time, and every rule carries the receipt that produced it: three drafts, three edits, one yes. There is no session where someone sits down to train anything. The training was happening already, in the five seconds it took to delete a sentence. It just finally has an audience.

The fastest way to stop fixing something is to notice you keep fixing it. Your edits are instructions. They deserve to be heard on the third time, not the thirtieth.