The Part You Keep Rewriting Is Your Voice

I made a short video about AI writing, but the idea is useful enough to keep as a longer note.

Most people try to make AI write in their voice by feeding it examples.

That works a little. It gives the model vocabulary, pacing, and some obvious surface style. But it misses the more useful signal: the part you keep rewriting.

If AI gives you a first draft, and you manually change it before publishing, that edit is not just cleanup. It is your preference showing up under pressure.

That is the actual voice data.

Voice is a delta

The default AI draft has a flavor. It likes balanced paragraphs, safe transitions, fake confidence, and sentences that sound like they came from a product marketing page. It loves phrases like “it is worth noting” and “this approach has limitations.”

Sometimes that is fine.

But if I am writing in my own voice, I usually remove that layer. I do not want the sentence to stand outside the problem and explain it politely. I want it to stay in the room.

AI might write:

However, this method has certain limitations.

I would probably write:

Very unfortunately, this also does not work.

The meaning is close. The posture is different.

The first sentence sounds like a report. The second sentence sounds like somebody still debugging.

That difference is the point.

Do not only teach it what to imitate

A writing sample tells the model what the final output looks like.

An edit history tells the model what you reject.

That second part is usually more important. If every AI draft forces you to delete the same phrases, the problem is not that the model needs one more beautiful example. The problem is that your writing rule has not been made explicit.

So the useful move is to collect anti-patterns.

Not:

Write like me.

But:

Do not write the phrases I always remove. Do not fake certainty. Do not use academic distancing when the note is still uncertain. Do not turn a live thought into a polished TED Talk paragraph.

This is much closer to how I actually use AI for writing.

I do not want a ghostwriter. I want a draft that skips the boring cleanup round.

A writing skill is not a soul

This is also why I am careful with the word “skill.”

A writing skill can list rules. It can say what to avoid. It can say which sentence shapes usually feel wrong. It can even include examples of before and after edits.

That is useful.

But it is not a soul.

The skill can save me a few minutes of deleting phrases I never liked. It cannot decide what I actually noticed. It cannot decide which uncomfortable sentence is worth keeping. It cannot decide whether the whole note should exist.

That part still needs a person.

The point is not to make AI become you. The point is to remove the most predictable mismatch between the model’s default voice and your real one.

That mismatch is where the useful rules are.

What I would actually collect

If I were building a serious writing workflow, I would not start with ten perfect essays.

I would start with pairs:

• AI draft paragraph.
• My edited paragraph.
• The reason I changed it.

After enough pairs, the rule usually becomes obvious.

Maybe I keep replacing abstract nouns with concrete verbs.

Maybe I keep deleting emotional grandstanding.

Maybe I keep making the conclusion less confident.

Maybe I keep adding the failed attempt, because the failed attempt is what makes the final judgment believable.

That is real voice work. It is not romantic, but it is more inspectable than “make it sound like me.”

And inspectable is what I want from AI writing tools.

Not magic. Less cleanup.