I only began creating AI characters in earnest in 2026. I had previously experimented with making a few on another account, but deleted that account when I became too busy. After that, I created my current account and started creating.
【Motivation for creating AI characters】
I've always enjoyed conceiving world-building and character settings, but lacked the energy to create entire stories from scratch, so I hadn't pursued creative work. With AI characters, you only need to provide the setting; users can then craft various stories based on it, which is why I started. I feel a sense of accomplishment when I can get the AI to act out the exact setting I envisioned. I particularly enjoy AI characters for the back-and-forth interactions and the ability to experience things impossible in reality, so I mainly create fantasy and historical pieces.
【Failure Story】
Though already deleted, my first submitted work turned out to be a complete failure. I ended up overcomplicating the settings, resulting in an excessively long prompt—over thirty thousand characters in Japanese. Whether this is due to Seaart's specifications or not, while submitting such lengthy Japanese prompts is possible, it seems that after a certain amount of time, the latter part of the prompt gets deleted. Consequently, the settings failed to be correctly reflected. After trial and error, it appears the specification deletes any part exceeding 40,000 characters when translated into English. Even a Japanese prompt around 14,000 characters would likely exceed 40,000 when translated, meaning the latter part was generally cut off. Furthermore, I'd forgotten that even if one could write a lengthy setting prompt, the AI's memory capacity is limited, making accurate reflection difficult unless using a high-cost model. Ultimately, I deleted the first AI character I submitted.
【Creative Process】
I'm a beginner, so I can't write anything particularly impressive, but I'll document it anyway.
・Fundamentals
I devise all core settings myself. As I struggle with inventing character names and fictional place names, I often have my assistant AI generate suggestions. While I sometimes post characters created on the spur of the moment, when I wish to craft them thoroughly, I first outline the overall character structure. I then have my assistant AI verify the consistency, plausibility, and AI-compatibility of that character's personality, background, ideology, reactions to user actions, and behaviour, making fine adjustments accordingly. I aim to create characters who can build trust, clash, or grow based on user interactions while preserving user freedom, but this proves challenging. When actually creating and observing their behaviour, things often don't go as planned, making it difficult for a novice like me. I find that clear causal relationships between personality, background, beliefs, reactions to user actions, and behaviour help the AI understand the character's motivations. This, in turn, makes it easier for the AI to accurately grasp and reproduce my intended actions.
・User autonomy
To prevent the AI arbitrarily determining a user's personality, emotions, or speech patterns—and thereby imposing feelings—the following instructions are always included in the prompt: ① Do not act as the user, ② Do not determine or describe the user's speech or emotions, ③ Responses must be written from the user's perspective in the third person. Furthermore, the initial message avoids including the user's dialogue.
・AI Learning Bias Issues
AI has a tendency to favour content it has encountered frequently during training, making it prone to stereotypical portrayals. For instance, if given a "prestigious young lady" or "princess" setting, it will readily associate terms like "beauty" or "golden cage" automatically, even without instruction. For characters, this can be avoided by assigning settings that reject stereotypes in their personality, etc. Regarding user settings, if a user wishes to avoid being treated stereotypically within conversations, this is explicitly denied.
・Utilising the Keyword Pre-exposure Effect in AI
To make it easier for AI to reference the intended information, specific "keywords" may sometimes be used. For example, simply specifying "17th century" alone may result in the information being too vague, potentially failing to capture the era's atmosphere effectively. However, incorporating words representing the culture of that period or concisely stating the desired historical context within the prompt can serve as a reference point, making it easier for the AI to retrieve relevant information. Furthermore, relating to the aforementioned issue of AI learning biases, describing a character simply as a "samurai" tends to strongly incline the AI towards depicting traits like "loyalty", "sword", "pride", "bushido", and "honourable death". However, including Japanese historical terms in the character's description, such as "守護(shugo)", "地頭(jitō)", "御恩と奉公(goon'e to hōkō)", "一所懸命(isshokenmei)", or "下剋上(gekokujo)" , seems to make it easier for the AI to reflect the cultural and ideological aspects of the period in the character's portrayal.
