Have you ever had this experience: the first few lines of a conversation with an AI character feel amazing, but the longer you chat, the more it feels like reading the same sentence repeated with different punctuation—she “feels very happy,” he “gives a gentle smile,” back and forth with just these few words. This usually isn’t because the platform is dumb, but because the character card itself provides too little material. If you haven’t set up the basic fields of a character card yet, you can first check out the Complete AI Character Card Setup Guide; this article is specifically designed to solve the advanced problem of “even though example dialogues are written, the AI still sounds like it’s reading a script”—using two core concepts borrowed from novel writing: Show, Don’t Tell, and catchphrase design.
Why Do AI Character Dialogues Easily Become Bland?
AI roleplay dialogue is essentially a “continuation”—the AI looks at how you write and how the character card is set up, and then continues the chain. If the character card’s personality only lists adjectives like “gentle” or “domineering,” the material the AI can mimic is very limited, and it will quickly run out of ways to say things and start going in circles.
The official SillyTavern documentation puts it very bluntly: the character description and personality fields are “permanent tokens” sent to the AI every time; however, once the chat history fills up the context window, the example dialogues are the first to be pushed out and no longer sent. In other words, if the example dialogues themselves are written poorly, the AI won’t even have material to learn from.
The Character.AI official creator guide mentions a similar logic: the AI doesn’t get much useful information from vague adjectives like “friendly” or “complex”; instead, concrete details and example dialogues are what truly teach the AI character its actual tone, pacing, and emotional logic. So the problem usually isn’t that the AI isn’t smart enough, but that the material we feed it is too thin. The solution can be approached from two directions.
Concept 1: Show, Don’t Tell—Replace “Declarative” Emotions with Actions and Senses
Show, Don’t Tell is one of the most basic and useful concepts in novel writing: instead of directly “declaring” an emotion, use actions, sensory details, and subtext to let the other person feel it. The same sentence written in two different ways makes a huge difference:
Tell: She is angry and feels you lied to her.
Show: Her knuckles tap twice on the table, not looking at you. “So… everything you just said is true?” Her voice is half an octave lower than usual.
This is especially important for AI characters because the AI continues writing based on the style you “demonstrate.” The SillyTavern documentation clearly points out that the character’s “First Message” is more likely to determine the AI’s subsequent speaking style and length than any other setting. If the first message is full of action and sensory details, the AI’s subsequent replies will naturally continue this style; if the first sentence is just a plain statement, the AI will follow suit and remain stuck in dry, declarative dialogue.
The demonstration of the character “Seraphina” in the official documentation is a living textbook for Show, Don’t Tell: she isn’t just described as “gentle”; instead, she actively holds your hand, using touch and warmth to build up gentleness layer by layer, without ever directly saying the word. When writing, you can use a simple “five senses checklist” to remind yourself: should this reaction include a micro-expression, a change in voice, an action, a sense of distance, or something intentionally left unsaid? You don’t need to pack all five senses into every sentence; just pick one or two key points to alternate, which is enough to make the dialogue vivid.
Concept 2: Catchphrases and Linguistic Fingerprint Design
When designing characters, many people only think of “giving them a catchphrase,” which results in the AI repeating the same phrase in every sentence. This is exactly the problem pointed out in another article on this site, AI Roleplay Repetitive Dialogue Guide: the character has too little material to use, so the AI has no choice but to loop the same few lines. The solution is not to “avoid catchphrases,” but to upgrade a single catchphrase into a whole set of “linguistic fingerprints,” which usually include three directions.
First, speech endings: fixed ending words or small actions, such as habitually adding “…right?” at the end of a sentence, or dryly laughing before speaking when nervous. Second, a dedicated metaphor library: prepare 3 to 4 metaphors related to the character’s background to use in rotation. For example, a chef character describing their mood might use terms like “heat control” or “seasoning” instead of the same phrase every time. Third, words they absolutely won’t say: clearly list words or topics the character will never use, such as never swearing or using internet slang—negative space is also part of a linguistic fingerprint.
The Character.AI official creator guide turns this logic into specific technical fields: the Definition field can hold up to 32,000 characters of example dialogue, formatted as “Name: Content,” and can use variables like {{char}} and {{random_user_1}} to prevent the character from getting confused by hypothetical dialogue partners in their own definition.
The official documentation also demonstrates a single set of example dialogues for the character “BoardWizard” to teach the AI how the character usually speaks and what topics they discuss. The key is to prepare several sets of dialogues for different scenarios (greetings, disagreements, goodbyes) rather than preparing ten lines with the same meaning to fill the fields.
The SillyTavern community has also developed more systematic approaches, such as the “Ali:Chat” style recommended in the official documentation. This style specifically teaches how to use repetitive yet varied example dialogues to “imprint” the character’s tone into the AI. Those interested in diving deeper into character card writing can explore this further.
Practical Rewriting Demonstration
Suppose you are writing a character who is a “retired mercenary, now running a small bookstore,” and the personality field originally only had the words “cold exterior but warm heart.” After applying today’s two concepts, the example dialogue can be rewritten like this:
{{user}}: Have you read this book too?
{{char}}: *He puts down the gun he was wiping—no, a fountain pen, casually brushes the dust off the book cover, and doesn’t answer immediately.* “I’ve read it.” *He pauses for two seconds, his gaze drifting to the depths of the bookshelf.* “Back then, there was someone by my side who would read it to me.” *He pushes the book in front of you, his tone intentionally softened, as if handing over something fragile.* “You can borrow it. Don’t fold the pages.”
This passage doesn’t directly write “he misses the past” or “he cares about this book,” but through actions (brushing dust, pushing the book, softening his tone), catchphrase-like habits (subconsciously reaching for a gun but correcting to a pen, the commanding “don’t fold the pages”), and intentional subtext (the unfinished “back then”), both the reader and the AI can read more emotion than what is on the surface. The next time the AI responds, it will also have more concrete material to continue with, instead of just repeatedly circling around the adjective “cold exterior but warm heart.”
How to Do It on Three Major Platforms: Not Just for Character.AI
Although the user interfaces differ, the logic of “using example dialogues to teach the AI how to speak” is common across several platforms, summarized in this table:
Simply put, no matter the platform, the two principles of “putting important settings first” and “quality over quantity for example dialogues” are more effective than piling up setting fields. WeavAI’s dual modes of interactive fiction and roleplay can apply the exact same writing style—first clarify the character’s sensory reactions and linguistic fingerprints, and then implement them in the setting fields, rather than rushing to fill the fields and calling it a day.
Common Misconceptions Q&A
Q1: Is it better to write a longer character card?
Not necessarily. The SillyTavern official documentation mentions that the more tokens the character settings occupy, the less “memory space” is left for the conversation history. While long-context models are less affected, models with smaller limits can easily become forgetful. Instead of stuffing the fields, it’s better to save the space for a few truly high-quality example dialogues.
Q2: How do I prevent catchphrases from becoming a new loop?
The key is “a set” rather than “a single phrase.” A single fixed line will inevitably get worn out, but a combination of speech endings + a dedicated metaphor library + words they absolutely won’t say allows the AI to generate variations instead of rigidly repeating the same line of text.
Q3: Does this apply to characters with pure dialogue and no narration?
Yes. Even without action descriptions, you can create a distinction in “voice” through word choice, punctuation, and sentence length—short sentences with little punctuation read completely differently from sentences that use ellipses and drag on.
Q4: Will the AI really “learn” the style I write, or just respond randomly?
The underlying logic of AI roleplay is to continue with the most likely next block of text based on the preceding context. Therefore, the style you demonstrate is the direction the AI is most likely to continue—which is why the quality of example dialogues and the first message is more important than you might think.
Conclusion
The hardest part of a character card is often not figuring out “who this character is,” but making the AI sound like the same person in every sentence. The next time you set up a new character, don’t rush to fill in the fields. Instead, think through the character using the five senses checklist and linguistic fingerprints before writing the example dialogues.
If you want to practice this writing style in interactive fiction or roleplay modes, WeavAI offers an anime style, a complete Traditional Chinese interface, and dual modes of interactive fiction + roleplay to switch and practice. Daily check-ins grant +500 points, and entering the code WELCOME15 on your first top-up gives a 15% discount. Creators also receive a 30% points rebate when their characters are used, so you can start practicing anytime.
