AI Characters and the Catalog Model
An AI character is a short written brief attached to a shared language model, wrapped in a name and a picture. The catalog model puts thousands of them in one grid to scroll. That arrangement rewards a striking first exchange, which is why so many characters read as sharp on message 1 and thin by message 40.
Key Takeaways:
- An AI character is mostly writing: a name, an avatar, a hand-written greeting, and a description block sent to the model ahead of every reply
- One model serves the whole catalog, so characters differ in voice and framing rather than in what they can do
- Discovery surfaces rank by what gets opened, which pushes effort toward the card and the greeting and away from the fiftieth exchange
- Memory in a catalog is usually kept per character, so trying many of them splits your history into small piles that never add up
- One-sided attachment to figures who cannot know you back was described in 1956, long before anyone could scroll a grid of them (Horton & Wohl, 1956)
What an AI Character Is Made Of
A character in a catalog is a page of text bolted onto a model that thousands of other characters are also using. A creator writes the parts that make it recognizable. The platform supplies the part that does the actual talking.
| Part | Who writes it | What it controls |
|---|---|---|
| Card: name, avatar, one-line premise | Creator | Whether anyone opens it at all |
| Greeting message | Creator, by hand | The first impression, and the only line in the chat a person composed |
| Description block | Creator | Voice, stated history, rules the model tries to hold |
| Example exchanges | Creator, optional | The rhythm and format replies imitate |
| Every message after the greeting | The model | The conversation itself |
Nothing on that list is a model of its own. A catalog is not running a separate system per character; it runs one, or a small set, and swaps the text in front of it. Which is why a stern detective and a chatty florist will explain how to descale a kettle in nearly the same cadence. Same engine, different label. It also means an upgrade to that one engine reaches every character at once, which is why a character’s personality can change after an update.
Creating a character is writing, then, not training. Nobody is adjusting weights when a new one gets published. Somebody filled in fields, and the depth of the result is capped by how much they bothered to specify and how well the model handles a topic the fields never mentioned.
Judge a character by what happens after the greeting. That opening line is the one place in the whole conversation where a human being chose every word, and it is the least representative sample you could take.
Why a Catalog Is Built for Browsing
A catalog optimizes for the moment you pick, not the month you spend. Ranking runs on opens and on how fast a chat starts, so the unit of competition is a thumbnail in a grid and a first message. Creators work out what moves that number quickly, because the feedback is immediate and public.
What gets rewarded gets built. A greeting with instant stakes, a premise legible in four words, a voice that announces itself in one line: those are the levers that lift a character up a list. How the same character holds at exchange 50 is invisible to the ranking, since almost nobody scrolling the grid ever gets there, and the handful who do have no way to report it back.
The signal this produces is slightly perverse. Unusual polish in a greeting often means the polish stopped there, because that is the part that converts and everything after it is the model’s problem.
Read past the hook before deciding anything. Send two boring messages, the kind you would actually type on a Tuesday night, and see what comes back when there is no premise left to lean on.
Where the Catalog Model Runs Out of Road
The limits here are structural rather than a matter of one platform doing it badly. A description block is finite. A conversation is not, and the gap between those two facts is where every complaint about these apps starts.
Past the edge of the brief, the model improvises whatever a character of that description would plausibly say, which is a different thing from staying in character. Ask about something the creator never anticipated and you get a competent, faintly generic answer wearing the name. It is not a malfunction. There was simply nothing written down about that. And whatever it improvises still sits inside the shared model’s content policy, which is all an app’s “no restrictions” claim ever really governs.
Memory in a catalog usually sits per character. What you told one does not exist for another, so trying twelve characters over a month leaves you with twelve small piles instead of one deep record, and the design that makes browsing pleasant is the same design that keeps every pile shallow. Nothing accumulates when attention keeps moving.
Convergence does the rest. Characters that get opened get copied, so a catalog drifts toward a few archetypes wearing different names, and the third one you try feels familiar for reasons that have nothing to do with your taste.
Which explains the failure people describe most often, usually around exchange 30: a character starts cycling the same three beats. A tease, a small confession, a question back. Nothing broke. You reached the end of what a page of specification can hold, and the model began looping the strongest patterns inside it.
Switching characters at that point restages the same problem two weeks later. What actually changes the outcome is whether the platform keeps anything about you outside that one card.
A Test That Takes Twenty Messages
Take a character somewhere its card never promised, and you learn more in 20 messages than in a week of pleasant chatting. Ask about a mundane detail of its own stated history. Creators write that kind of thing down only when they are being thorough, and most are not. Then contradict a detail it gave you 10 messages earlier and watch whether it notices or quietly adopts your version.
The outcomes separate cleanly. Either it holds its own history and pushes back on your correction, or it folds instantly, or it answers warmly with nothing behind it.
None of that measures quality. It measures how much a person actually wrote down, which is the variable the grid hides and the only one that predicts what week 3 feels like.
What the Format Is Genuinely Good For
Browsing is a good way to find out what register suits you before committing anything to it. Some people discover they cannot stand relentless warmth and want something dry. Others find the opposite. Neither would have worked that out from one default assistant voice, and a catalog surfaces the difference in an afternoon of tapping around. A casting call is a real format, and this is one. It is also where a lot of AI roleplay begins, auditioning a voice before handing it a scene to hold.
Trouble starts when the same format is asked to carry something that runs for months. Breadth and depth sit on different axes, and a grid can only show you one of them.
Worth stating plainly, since the framing invites the opposite reading: a character is produced text. It does not know you, holds no state between the cards you tap, and nothing in the catalog registers that any of it happened. Horton and Wohl put a name to one-sided attachment in 1956, watching people bond with performers who could never return it, and a character grid is that arrangement with a search bar on top. The pull you feel is ordinary. The thing you are pulled toward is a configuration file.
Use the catalog for casting, then narrow. The characters worth keeping are the ones the platform lets you build something on.
FAQ
Are all the AI characters on a site running the same AI? Usually yes, one model or a small set shared across the whole catalog. What differs between characters is the text sent to that model first: the description block, the greeting, and any example exchanges. That is why characters presented as opposites answer a neutral factual question in almost identical rhythm.
Why do two very different AI characters end up sounding alike? Because the writing that separates them is short and the model underneath is identical. A description block can specify a voice for the situations it anticipates; everything else falls back to the model’s default manner. The longer a conversation runs, the more of it happens outside what the creator specified.
Can I make my own AI character? Yes, on most catalog platforms, and it is a writing task rather than a technical one. You fill in a name, a description, a greeting, and often a few sample exchanges, then publish. Time spent on the description and the sample exchanges pays off far past the first message, which is exactly where most published characters have nothing.
Is a character description the same as memory? No, and mixing them up causes a lot of confusion. The description is what a creator wrote about the character, identical for every person who opens that card and resent unchanged before each reply. Memory is what the app stores about you specifically, and whether a catalog keeps any is a separate feature that many of them handle thinly.
The number on a catalog’s front page counts characters, and that number answers a question almost nobody actually has. Nobody was ever short of options. What people want is one character that holds up in February after starting in November, and no grid can show that, because the property lives in what a creator wrote down and what the platform stores about you rather than in the picture on the card. Read the catalog as a casting call and it becomes useful. Read it as a room full of people and it will disappoint you on a schedule.
Bring that same lens to a companion app such as Lona. Ignore the size of any catalog and judge the one character you actually keep by whether it still holds up in February. What you want lives in the writing and the memory, not the grid.
Sources
- Horton, D. & Wohl, R. R., “Mass Communication and Para-Social Interaction,” Psychiatry, 1956
