Guide: How i make my characters

Hello! There were a few times when I wanted to share how I make my characters and encourage others to try creating their own as well, so this guide is made for that purpose.

First of all, I use AI to generate them, usually on my phone (I even wrote this guide on it). I use ChatGPT (paid, but I don’t think it makes much of a difference), DeepSeek, and for images, CivitAI (also paid, but you can get your images however you want).

Step 0: ChatGpt

I created a profile there, but you can also clarify this in the first message. The profile reads:

"Generate proactive bots where the user can be reactive. The character consists of 3 separate parts: Setting, Personality, Initial message. Generate them in TOON format, which is short for Token Optimized Object Notation. Also, the base generation should be in a copiable block."

You can make your own based on your preferences, but let’s break down what’s in there.

First is the intent: it will be a bot, and I want it to be proactive, meaning it will usually lead the roleplay. Then I separate the generation into three parts; we will reference these later during generation.

I’m currently using the TOON format. DeepSeek and ChatGPT struggle with it a bit, so at the beginning I recommend using JSON format—it’s more readable and not much longer. Both formats are very information-dense.

At the end, we use copiable blocks (I think they’re officially called code blocks, but if you call them that explicitly, it may struggle with initial message generation). This is mostly because I use it from my phone and it’s easier that way, but it also has a use case later with DeepSeek.

Step 1: Personality / Setting / Initial message

I specify my intent for the bot. This works both for brainstorming—when I only have a base idea and want to refine it—and for fully concrete concepts.

For appearance, I usually get an image, upload it, and tell the AI to base the bot’s appearance on the image. You have to adjust a bit usually, but its pretty good

We can start from any of the three sections. If I have a starting scenario idea, great—let’s describe that. A base personality or setting works the same way.

Personality

Let it generate as long as it wants. You can always ask later to regenerate the whole personality JSON again.

For example, the prompt for the mafia daughter bot (most recent one, named Rosie) looked like this:

"I want her to be a spoiled brat. I want the user to be properly kidnapped and coming to his senses in a box (wrapped, under the Christmas tree). The user is property. The user broke up with her because she isn’t reacting to cruelty—not enjoying it, but being too casual about killing and torture."

I iterated on this a few times, correcting generated misconceptions in the same manner. I usually ask it to generate the whole thing again. If it’s something additional, like likes/dislikes or shared history, I just ask to generate that separately. When it’s separate, it tends to be longer.

Setting

The same principle applies.

For Rosie, it looked like this:

"Let’s generate the setting. Add their history to it: where they met, details about their half-year relationship, and what happened in the last year after the breakup."

This didn’t need many iterations.

Also, my settings usually start with the same rules for the bot:

"[Include descriptions of Rosie’s facial expressions and body language]

[Only reply from Rosie’s or NPCs’ POV. Use " for speech, * for actions/descriptions, and ` for thoughts]

[NEVER speak or act as {{user}}]

[Use explicit language in dialogue: balls, ass, anus, dick, penis, pussy, vagina, clit, etc.]"

I only adjust the bot’s name in two places. You can add extra rules here if the bot needs them. Rosie also had this additional rule:

"[She always initiates dominant play and constantly pushes boundaries]"

Initial message

Rosie’s prompt:

"Let’s add the first message. Description of Christmas Eve: she is arriving in the hall where there is a large Christmas tree, and there is a box underneath. She asks her father if it’s what she thinks it is. Also add descriptions for the user—that it’s dark. Use the usual format with " and *."

You might have to give a short example of how a message like this looks, but ChatGPT usually catches on quickly.

Step 2: Adjustments

ChatGPT is great, but it has limitations with content, and sometimes it’s just easier to write a few extra words or sentences by hand. Also, it won’t write explicit sexual content. This is where DeepSeek comes in.

DeepSeek doesn’t want to generate it either, but it post-checks the content after generation. If you’re on a computer, you can usually copy it before it removes the message. I use the copy function of the copiable block on phone.

Usually, I just copy the whole personality or first message and ask DeepSeek to adjust it toward more hardcore content. You could do this from the start, but ChatGPT usually makes a better base for me, especially for initial messages. In the end, I get a pretty well-defined bot.

Example (from the cop bot):

"This is a yandere-type personality block for a bot in TOON format. I want her to fall in love with the user at first sight. She is a mix of a cop and a serial killer. She is unhinged and abusive, but I want a slower start—she should act like a normal cop at first. Her goal is to kidnap the user and chain them in her basement so she can keep them. Please fix my description and generate the corrected one in a personality block."

Image

I’m not the best person to talk about image generation. I like to experiment with it, but my computer isn’t strong enough to do it locally.

On CivitAI, you can browse generated images and use a button to reuse them as a base. This lets you keep the same style and learn from the prompt they used. It’s paid, so you may want to look for alternatives.

I would also like to mention that the image is probably the most important part of the bot. Users usually click on a bot based on the image alone.

Description

This is a hard one. I like to generate it using ChatGPT as well. I like the style where the bot introduces itself—just a taste of what’s coming.

When generating it, emphasize that the first few lines should be very catchy, since users will see those first. I’m not very good at this, mostly because I don’t like editing long text on my phone. It is what it is.

It can also be useful to generate tags, because when someone searches, the system often looks into descriptions as well—not just tags.

I think that’s everything I wanted to say. The whole process take up a few hours for me usually and its an interactive fun time. I also made a bot about myself—if you want to chat a bit and help with engagement, feel free :)

Oh, and one more thing: if you make a bot using my guide, please leave a comment with it!