Atlas AI Resources

Article

My Grok Bots Learn My Voice From What I Actually Post

By Austin Hoopes · Published November 16, 2026

I can tell a bot how I want to sound, and it helps a little. What helps more is the bot finding out what I actually posted.

Several of my Grok Bots draft things that go out under my name. This piece is about two of them: the one that drafts short X posts, and the one that drafts long pieces like this one. Each keeps a small file of voice notes, and the rule for that file is narrow: notes get added after comparing the drafts with what I actually published. Here’s what’s in them, and the part of the loop I keep leaving open. About a four-minute read.

Describing my own voice is a guess

If you asked me how I write, I’d say plainly, and not too long. That’s true. It also doesn’t tell a bot much about which of three drafts I’ll pick on a given day.

What I post isn’t a guess. It went out, with my name on it. So that’s what the notes are built from.

The official Grok Bot guides point the same way. One describes sending the difference between a bot’s draft and what the person actually sent back to the drafting bot as feedback, so the next draft sounds more like them. (The summary is in Grok Bot for post-sales.)

What the notes actually say

The X post bot’s file is four lines long right now. Each line is one observation, with the reason it’s there:

None of that is a style guide in the usual sense. It’s closer to a scorecard. Two lines say what worked, one says what didn’t, and one marks an exception.

Numbered options make the comparison easy

Every draft I get to copy comes as numbered options, each alone in its own code block. (That rule is in The Standing Rules My Grok Bots Share.) I asked for that so I could copy with one tap.

It has a side effect I like. When I post Option 3 word for word, there’s no question which one I picked or whether I changed it. The note can just say so.

Untouched is the strongest signal

A correction tells a bot what was wrong. A draft I post untouched tells it what was right, in exactly the words I was willing to put my name on. Two of the four notes are that kind.

Silence counts too. The retired phrase didn’t fail loudly. Nobody picked up on it. That’s worth writing down, because otherwise a bot that liked the line would keep reaching for it.

The gap in my loop

All of this only works if the bot finds out what went out. That’s the weak part of mine.

The bot that drafts my long pieces also keeps a ledger of follow-up quote-post moments for older articles. Each row has a status: unused, drafted, posting unconfirmed, or used. Used means I confirmed I posted it. As I write this, every row that’s been drafted still says posting unconfirmed.

That’s honest bookkeeping. It’s also a bot that can’t learn anything from those drafts, because nobody told it what happened to them.

Its voice file shows the same thing. It has a heading and one line explaining that notes get added after comparing drafts with what I published. No notes yet.

So one bot has a scorecard and the other has a heading. The difference isn’t which bot is smarter. It’s whether the result makes it back.

Why a file and not just memory

The official guides say Grok Bot’s memory can learn your writing style. I don’t doubt it. I still want these notes in a plain file I can open.

It’s the same reason as everything else in my setup that has to be exact: I can read it, check it, and delete a line that’s wrong. (That rule, and a few others, are in What Running Thirty Grok Bots Actually Looks Like. For how bots keep what they know more generally, see Agent memory for AI fleets.)

To be clear about what this is: it’s not a Grok Bot feature. It’s a short text file per bot and a habit of comparing drafts with what went out.

If you want your bots to sound more like you

If you’d rather start from your own sent mail and messages, the prompt library has four prompts for a personal style guide.

Which of your bots is still guessing because nobody told it what happened to its last draft?

About the author

Austin Hoopes runs a fleet of Grok Bots that handle his inbox, research, and writing, and writes here about what actually holds up day to day.

← All articles