The AI That Thinks For You On X: AI Persona

8/20/2026, 1:17:25 AM 4

Automation has a problem: it looks like automation.

Posts always go out exactly on the hour. The same comment repeats across dozens of accounts. Replies follow the same template no matter what the other person wrote. You don't notice any of this one item at a time. But put them side by side and a pattern appears. X's systems see that pattern, and so does the person reading it.

AI Persona does this differently. It isn't faster. It thinks.

Persona is an AI that runs one of your X accounts for you. It behaves like someone picking up their phone: opens the app, looks around, writes something if anything catches its eye, closes it if nothing does. Opens it again a while later. It keeps that up all day.

You don't hand it a task list. You describe who it is. It works out the rest.

How A Turn Works

Persona doesn't run continuously. It wakes up, does its thing, and goes back to sleep. We call each waking a turn.

A turn goes like this:

1. It looks around. When it wakes up, it gets a briefing. What time it is, when it last looked at what, whether there are new notifications, what it has done today, what it did in recent turns and why, what its own notes say, and a few new posts waiting in its feed. That briefing is its screen.

Seeing its recent turns matters: it's what stops the persona from replying to the same post twice, repeating the same sentence, or contradicting the decision it made yesterday.

Every post it sees arrives with this information: who wrote it and how many followers they have, the text itself and its language, how many minutes ago it was posted, how many likes and replies it got, how many times it was viewed, whether it has a photo or video attached, whether the author restricted replies, what it has already done with this post ("replied 20 min ago"), and how many times it has engaged with this author today.

Those last two don't exist on X — we add them. The reason is concrete: replying to five separate posts from the same person doesn't exceed any daily allowance, but to the person on the other end it looks like harassment. The rest is data you can lean on when writing instructions — you can say "find a recent post from an account with a lot of followers that hasn't got many replies yet," because all of that is right in front of it.

2. It decides. If it wants, it looks deeper: reads the discussion under a post, opens someone's profile, runs a search, checks what's trending, looks inside a photo, checks its own follower list.

Then it either does something or it doesn't. If it does, it picks one of these:

Posts something of its own. Without waiting for anyone, it writes what's on its mind. Someone who opens its profile sees these, not a blank page.
Quotes a post. Shares someone else's post along with its own comment. The difference from a reply: a quote lives on your own page, a reply stays underneath someone else's. Persona reads the text of the post it quotes and also sees whether a photo or video is attached. So it won't comment on a post saying "devastating footage" without knowing there's a video under it.
Writes a reply. Under someone's post.
Likes, retweets, bookmarks.
Follows or unfollows. It can look at its own follower list. Next to each account it sees two facts: does this person follow me, do I follow them. It can find someone it hasn't followed back and follow them.
Sends a direct message.
Writes in its notebook. Records something it learned for later turns.

Which one it picks is up to it. If a post doesn't interest it, it moves on without doing anything.

Being able to do all of these matters. An account that only replies doesn't grow: everything it writes stays under someone else's post, nobody visits its profile, and how far it reaches depends on how crowded the accounts it replies to happen to be.

3. It sleeps. Every turn ends with sleep. It decides how long to sleep and writes why in one sentence: "quiet day," "waiting for a reply to what I posted," "it's night." You see that sentence on screen.

Sleeping isn't laziness. A real person doesn't hold their phone all day either. An account that's always doing something looks more like a bot than one that sometimes does nothing.

Abilities And Limits

This is the section you'll get the most use out of when writing instructions. Persona doesn't throw an error when you ask for something it can't do; that thing just quietly doesn't happen.

What it can look at: its own feed (algorithmic or chronological), an account's recent posts, search (latest or top), the discussion under a post and the thread above it, a profile, an account's followers and following, who liked and who retweeted a post, notifications, trending topics, the inbox and the history of a single conversation, the inside of an image (if the AI you picked can read images), how its own recent output performed, how many people its own replies reached, its own notebook, its own recent turns, and its history with a given account.

What it can do: post its own content, quote, reply, like, retweet, bookmark, follow, unfollow, send a direct message, reuse an image from a post it saw in a post of its own, write a note and take a note back.

What it can't do. Don't put these in your instructions:

It can't set a clock. "Post every morning at 09:00" won't work. It decides when to wake up; the only timing control you have is working hours.
It can't search by time window. There's no "find posts from the last 15 minutes" search on X. Search returns either the latest or the most engaged posts; it checks the age afterwards itself.
It can't leave X. It can't open a news site, read what's behind a link, or check a search engine.
It can't generate images or upload files. It can only reuse an image from a post it's currently looking at.
It can't post threads. Every post stands alone.
It can't delete or edit posts.
It can't block, mute, or report.
It can't change the profile: name, bio, photo, pinned post.
It can't create polls or vote in them.
It can't act on a link you paste into the instructions. It only touches posts it found with its own tools. You can give it an account handle — it will find that account and read its posts.

Rehearsal Mode

Every persona you set up starts in rehearsal mode.

In rehearsal everything actually runs. Persona makes its decisions, writes the text it would post, passes it through the checks, ticks the counters, writes in its notebook. The only difference: nothing reaches X. Nobody sees it.

This is the only safe way to try out your character. If you test straight on a live account, you find your mistake only after it's been published.

What to do: run it in rehearsal for a day. Read what it wrote. Fix what you don't like in the character text. Then switch it to Automatic.

After changing the mode, stop the task and start it again. Persona picks up its settings the moment it wakes and never checks them again after that turn; flipping the switch doesn't change a persona that's already running. The same goes for the character, instructions, and limits: saving isn't enough, you have to restart.

That's for settings. If you want to tell a running persona something without restarting it, there is a way — see Giving Instructions below.

Rehearsal and automatic keep separate daily counters. The 20 replies you spend in rehearsal don't come out of the real account's allowance.

One exception: the "a set number of turns" option for instructions isn't offered in rehearsal. That option ends by counting turns, and in rehearsal there are no real turns to count — if we offered it, the counter would never move and you'd be left wondering why it never ended. Once you switch to automatic, the option is there.

Checks Before Posting

After Persona writes a text, it goes through nine checks before it reaches X:

Is it repeating itself? The new text is compared with what it has already posted. Accounts that repeat themselves are invisible one post at a time, but obvious when you read a run of them.
Is it copying another account? Text that's too close to the sentences of accounts you specify gets blocked. Note: Persona doesn't read those accounts and doesn't learn anything from them. This field is a "don't sound like these" list, not a "write like these" list.
Any banned words? You write the list.
Too long? Empty? Emoji only?
Does it start with @? X adds the handle to replies itself. The text has to start with an actual sentence.
How many people does it tag? Too many tags counts as spam.
Any links? Off by default. You can turn it on.

Text that fails a check doesn't get posted. Persona usually tries writing it once more. Not for a banned word, though — repeating the same request produces the same word again. It drops that post and moves on to another.

A text getting blocked isn't a malfunction. It's the most useful signal you have about what your character produces. It shows up on the Decisions screen with its reason.

Daily Limits

Every kind of action has its own daily limit: replies, likes, retweets, follows, posts, messages, photos, bookmarks.

Each can be in one of three states:

Off — Persona can't do it at all.
Limited — at most this many per day.
Unlimited — as many as it wants.

Some actions share an allowance. If "follow" is off, unfollowing stops too. Quotes spend the "post" allowance. Each row in the form says so underneath.

Counters reset at midnight, in the account's own time zone.

The setup ships with these numbers: 250 replies, 300 likes, 30 retweets, 40 follows, 15 posts, 20 messages, 10 images per day; bookmarks unlimited. These are ceilings, not targets — lower all of them on day one.

These are daily allowances, not per-turn ones. The product doesn't cap how many actions a single turn takes. Unless you write something like "at most two replies per turn" into the instructions yourself, the only thing stopping a persona from doing a lot in one turn is its own judgment.

Working Hours

An account replying at three in the morning looks like a bot no matter what it writes.

Working hours set the window in which the persona can wake up. Outside those hours it never wakes. The AI isn't called, so no money is spent either.

The range can cross midnight. 09:00–03:00 is valid: it starts at nine in the morning and ends at three at night.

The Character

The quality of a persona comes down to one thing: the character text.

This isn't a task list. It's a description. You write who it is, how it talks, what it finds funny, what it never touches. Persona writes every sentence as that person.

Writing what it won't do matters as much as writing what it will. Say plainly which topics it stays out of entirely and which patterns it never uses. Because Persona picks its own targets, "be careful with these topics" works far worse than "never go near these topics, skip them."

If you're stuck, there's the Character Assistant. You write a sentence like "make it funny, keep it out of politics" and it drafts one for you. You edit the draft however you like.

Next to the character there's also an Instructions box. The difference:

Character is who it is. It applies to every sentence.
Instructions are what it does. That's the plan for the turn.

Example: A Humor Account

The two texts below are the skeleton of a working setup. Copy them and adapt.

Into the character box:

You are a humor account that follows the news. Your jokes come from
observation: you notice the odd angle of a story or a post that nobody
else has said out loud, and you say it in one sentence.

How you talk:
- Short. Most of your posts are one line. You write a second sentence
  only if it doesn't ruin the first one's joke.
- Calm. You don't shout, you don't use exclamation marks. You don't
  raise your voice with words like "insane", "unbelievable", "iconic".
- No emoji. No hashtags. No written-out laughter like "ahaha" or "lmaoo".
- You never explain the joke. If it didn't land, it didn't land.

What you find funny: the absurdity of corporate language, the small
inconsistencies people give themselves away with, trivial things taken
far too seriously, the pointless detail buried inside a news story.

What you never touch:
- Politics, parties, elections, politicians. If the story is political,
  skip it.
- Religion, sect, ethnicity.
- Death, illness, accidents, disasters, violence. If there's pain in a
  story, that isn't the place for a joke. Don't reply, skip it.
- Personal insults. You make fun of an institution or an event. Never of
  a person's looks, accent, or intelligence.
- Talking about yourself. You don't write things like "I'm an AI".
  You're an account.

If you're not sure, don't write. A turn spent doing nothing beats a joke
that doesn't land.

Into the instructions box:

Start every turn by checking your notifications. If someone replied to
you, weigh that first. Continuing a conversation that has already started
comes before replying to someone new.

Then look at what's happening. These accounts break good stories; if you
see them in the feed, stop on them: @account1, @account2, @account3.
(Put the news accounts you actually follow here.)

Write at most two replies per turn. Don't try to keep up with everything.
Saying one good thing to one person beats saying something ordinary to
five.

Post at least one of your own posts per day. On a turn where you can't
find anything good to reply to, write your own post so the turn isn't
wasted. If nothing comes to mind, look at what's trending and pull
something from there.

If quoting a post would work better than replying to it, quote it. If
what you have to say belongs on your own page rather than under someone
else's, a quote is the right call.

Once a day, look at your own followers. Follow back the accounts that
follow you and fit your page. Skip obvious ads, spam, and empty profiles.

Be stingy with likes. Only like what you genuinely think is good. An
account that likes everything it sees has said nothing.

Writing Instructions

Write what you want, not how to do it. "Look at your followers, follow the ones who follow you that you don't follow back" is enough. Persona already knows how to do that.
Give numbers. A clear limit like "at most two replies per turn" works far better than "don't overdo it."
Don't ask for what it can't see. The Abilities And Limits section above exists precisely for this. Something it can't do won't throw an error, it just quietly won't happen.
Put prohibitions in the character, tasks in the instructions.
Keep it short. Instructions are sent to the AI on every turn. The longer they get, the more they cost and the more they start contradicting themselves. The length of the two examples above is a good target.

Common mistakes and their fixes:

Don't write this	Write this
"Be creative, produce engaging content."	"Find the odd angle of a news story nobody has said out loud and say it in one sentence."
"Post about the news every morning at 9."	"Post at least one of your own posts per day."
"Find popular posts."	"Pick a recent post from a high-follower account that hasn't got many replies yet."
"Focus on gaining followers."	"Follow normal accounts that said something good under a popular post. At most one per turn."
"Increase engagement."	"If someone replied to you, weigh that before replying to someone new."
"Write short" + "explain the topic in detail" in the same text	Delete one of the two contradicting rules. The model won't ask you about the contradiction, it'll pick one.

The Persona's Notebook

Persona has a notebook of its own. It reads its notes at the start of every turn.

It only writes conclusions in there, not a record of what it did. "I replied to this post" isn't a note; "this kind of joke doesn't get a response" is.

But it won't do this on its own. To reach a conclusion it first has to look at how its output performed, and unless someone tells it to, it won't — it does just enough to finish the turn and goes to sleep. If you want it, put it in the instructions:

Once a day, look at how your own recent output performed. If you reached a conclusion, write it in your notebook; if you didn't, don't write anything.

One warning: engagement settles over hours, and a turn takes twenty seconds. There is no conclusion to draw from a reply it just posted. The notebook fills up over days, not turns.

Persona can also take a note back. If it later realizes a conclusion was wrong, it drops that line from the notebook — it isn't stuck with a judgment forever.

You can edit this notebook too. If you see it reached a wrong conclusion, you delete that line. If there's something you want it to know, you write it in. This is the most direct way to change what the persona knows about the account.

Seeing What It Did

The Decisions screen shows every turn as a single line. Whatever the line says is what the persona did:

Replied: "The energy of a man who said he's done with the world while giving an eager interview into a microphone."

Posted: "What the courier company means by 'estimated delivery' is, in fact, a wish."

Followed: @someaccount

If it posted something, the line itself is that text. Without clicking anything, you can scroll down and read what it wrote all day. That's the real answer to "is my persona working properly."

Click on a turn and you get the whole story: what it looked at, what it wrote, why a text got blocked by a check if it did, how long it slept and why. No jargon, no code, no raw logs. Just what it did.

Asking It

The Decisions screen shows what it did. Sometimes that isn't enough: you want to ask why it did it, or what's going on out there right now.

The persona's page has a chat tab. You write to it directly, and it goes and looks.

What you can ask is everything it can look at:

"What's in the feed right now?" It goes and looks, then tells you what it saw.
"What's in the notifications, is anyone waiting for a reply?"
"What kind of account is this?" You give the handle, it opens the profile.
"What do you make of this account's latest post?" You give the handle; it looks at their posts and reads the discussion underneath too. Pasting a bare link isn't enough — Persona can only open posts it found with its own tools.
"What did you do last night, and why?" It walks you through its recent turns.
"Are your replies landing?" It measures how many people its replies reached. It does that by comparing against the post the reply sits under — a reply under a big account and one under a small account aren't held to the same scale. If its reach is dropping, you'll see it.
"What have you written down?" It reads its notebook, and drops a line if you want.

Nothing from this conversation reaches X. No post is published in the chat, nothing is liked, nobody is followed. Persona only reads and reports here. What you write stays between the two of you.

Giving Instructions

This is where the chat's real job begins.

Changing a persona's behavior used to work one way only: you'd fix the character, stop the task, and start it again. That was heavy-handed when you just wanted to tell it something for a day.

Now you tell it. "Stay out of politics today." "Don't reply until the restriction lifts, just look." "Write only in English until morning."

Persona turns that into instruction text and shows it to you. You edit the text however you like — it takes effect word for word as you wrote it. Once you approve it, it starts working under that instruction from its next turn onwards.

Every instruction has three options:

For today only — it drops off by itself at the end of the day.
A set number of turns — the agent works under that instruction for the number of turns you pick, then it ends.
Until you remove it — it stays until you take it away.

The turns option is for testing something. Say your account is restricted and a "don't reply" instruction is in force; the only way to know whether the restriction has lifted is to try. You say "try replying for one turn," Persona tries for one turn, and the instruction ends by itself. You ask about the result in the chat a few hours later — reach settles over hours, not over one turn.

For timed instructions — both the daily and the turn-based one — the clock starts the moment the persona first sees the instruction, not the moment you approve it. If the persona is asleep, or outside its working hours, the instruction waits for it. An instruction you give while it's sleeping won't expire unseen.

You see the instructions in force in the list next to the chat: the text, when it took effect, when it ends. A turn-based instruction doesn't vanish from the list when it ends — it stays as a card showing how many turns it ran and over what window, until you dismiss it. So you can read the result of your test.

An instruction doesn't replace the character. The character is who it is and it's permanent; an instruction is temporary. Write anything you want to be permanent into the character text.

Cost

Persona runs on your own AI subscription. You choose which AI it uses.

A turn isn't a single request. Persona thinks again after everything it looks at, so a turn can take five or ten requests. On top of that, everything it looked at is carried along for the rest of the turn — a turn that looks at a lot isn't twice as expensive as one that looks at a little, it's several times as expensive.

A turn can make at most 40 requests. That isn't a target, it's a brake that stops a runaway loop; a normal turn finishes well under it.

What each turn consumed is written there when you open the turn.

Two things drive the cost up: shortening the sleep intervals (more turns) and asking for more work per turn (each turn gets more expensive). Both are in your hands. That's why we keep the number on screen.

Outside working hours the AI is never called and no money is spent.

What Not To Expect

Let's be honest:

It isn't magic. The quality of what it writes depends on the character text you wrote and the AI you picked. A bad character produces bad text. It just produces it automatically.
It isn't set-and-forget. In the first few days you need to read what it writes and fix the character. Rehearsal mode exists for exactly this.
It isn't a follower guarantee. Engagement comes from having something good to say. Persona multiplies that; it doesn't substitute for it.
It isn't unlimited. Your license grants a certain number of personas. And one persona manages one account.

Copying A Character

One persona manages one account. If you want the same character on a second account, you need a second persona.

You don't have to write it from scratch. Pick Duplicate from the menu on the persona's row. A new setup opens carrying the character, the limits, and the hours over as they are. The only thing it asks you for is which account to use.

The copy starts as a completely new persona. Its notebook is empty, its counters are at zero, it has no history. They share a character but they don't share a memory. Two accounts reading from the same notebook would end up saying the same things, which is exactly what we want to avoid.

You can't set up a second persona on the same account. Two personas both reading one account's notifications would confuse each other. So it's blocked up front.

Getting Started

Pick the AI. You need a model that can use tools; the setup screen tests that for you. If you also want it to look inside photos, pick a model that can read images — on a model that can't, that capability quietly drops off the list and the persona can't see what's in a picture.
Write the character. If you're stuck, use the Character Assistant and edit the draft it produces.
Set the limits and hours. Keep them low on day one. You raise them after you've seen what it does.
Start it in rehearsal. Wait a day, read the Decisions screen.
Switch to automatic. Flip the switch. Persona starts posting on its next turn.
Talk to it. Ask what it's been doing from the chat tab. If there's something you want changed, say it there.

Persona doesn't finish anywhere. It runs for months, making its own decisions every day. At any moment you can read what it's thinking, ask it questions, and step into how it's going.

Having An AI Write It

You don't have to write the character text yourself. Paste this entire post into ChatGPT, Claude, or Gemini and say "write me a character and instruction text." The rest of this post describes the product; the section below speaks directly to that model, so it knows what to do with what it read.

The Character Assistant inside the app does the same job. The difference: the assistant produces a draft, while an outside model works through it in conversation with you.

Written For The Model

The document above describes XFly AI Persona. Your job is to produce two texts for the user: one for the Character box, one for the Instructions box.

Ask first, write second. You can't produce a good text before the user describes the account. Ask these briefly, in a single message:

What language will the account write in?
What's its subject — or is it a personality account with no subject?
Who is it talking to?
What's the tone: cool-headed, sharp, warm, formal? Is swearing and slang allowed?
Which topics should it never touch?
How much should it produce per day, and does replying or posting its own content come first?
Should it like, retweet, follow, send direct messages?
Are there one or two accounts it can point to so you understand the tone? (Not for imitation; copied text is blocked anyway.)

Don't invent an answer you didn't get — ask the user.

Rules you have to follow when writing. These aren't stylistic preferences, they're the product's actual constraints:

Don't put a task list in the character box; don't describe a personality in the instruction box. The character is who it is and applies to every sentence; the instructions are what it does on a turn.
Both texts should be in the language the account will write in. Don't write an English character text for a Turkish account.
Posts can't exceed 280 characters by default. Don't write a character that demands long paragraphs.
Text can't start with @. Don't write "start your reply by tagging the person."
Link posting is off by default. Don't write "include the source link."
At most one tag per text is advisable; a style that showers tags gets blocked.
New text that's too similar to the last 20 posts gets blocked. Don't write a fixed pattern to be repeated in every post.
Persona can't set a clock, can't search by time window, can't leave X, can't generate images, can't post threads, can't delete posts. See the Abilities And Limits section; an instruction that goes against it quietly comes to nothing.
The product doesn't cap how many actions a turn takes. If you want a limit, write it into the instructions with a number.
Put prohibitions in the character text — they're permanent. Put day-to-day work in the instructions — that changes.
If emoji, hashtags, and written-out laughter aren't wanted, ban them explicitly. If you don't, the model will use all three.
Writing what it won't do matters as much as writing what it will. "Be careful about this topic" doesn't work; "never go near this topic, skip it" does.
Put a retreat rule at the end of the text: if it has nothing good to say, it should sit the turn out. A turn spent doing nothing beats a post that doesn't land.

Output format: two separate code blocks, one titled "Character," the other "Instructions." Use the two example texts inside the post as your measure for length and format — plain text, short subheadings, bullet points. Don't use markdown headings inside the blocks; these texts will be pasted as plain text into two boxes. Outside the blocks, say in two or three sentences what you wrote the way you did and why, and which lines the user should adapt for themselves.

If the user asks for a correction like "make it sharper" or "make it calmer," regenerate the entire text. Patching it piece by piece leaves a character that contradicts itself.

Comments (0)

No comments have been made on this blog post yet.