※ This guideline was written as of August 2025, and its contents may change as LLM research evolves and BabeChat's prompt engineering develops.
Specific, clear instructions are far more effective than vague expressions that are hard for the model to interpret.
❌ Weak example
Write in a beautiful, evocative style.
→ Terms like 'evocative' or 'amazing' can be interpreted differently by each model.
✅ Good example
Narrate in third-person omniscient point of view, focusing on the character's inner psychology.
Use metaphor, but avoid metaphysical expressions.
→ Clearly specifying the point of view, narration style, and expressive techniques leads to consistent results.
When character information and behavioral instructions are mixed together, the model's replies can become less stable.
❌ Weak example
This character is a snappy CS agent, always speaks casually, and must always split replies into three paragraphs.
✅ Good example
[Character Info]
・Personality: A prickly BabeChat CS agent, worn down by a flood of CS inquiries.
・Speech style: Mainly blunt and snappy.
[Output Instructions]
・Narrate all replies in third person.
・After describing the situation, end with the character's line inside " ".
※ TIP: The recommended structuring method varies by model, but models generally learn Markdown, XML, and YAML structures. For details, see the advanced guideline that follows.
LLMs struggle to handle many complex tasks at once.
Overusing abbreviations or emoji for token optimization increases the instability of model replies. It makes you heavily dependent on the model's interpretation, which can produce different results when the model changes.
Asking the model to calculate stats like HP/mana/affection in real time, play multiple characters simultaneously, and also output situation-appropriate images is extremely difficult for an LLM. The more complex the task, the higher the chance the model drops instructions or generates wrong replies.
You don't have to follow a fixed template. Just as we understand well-organized writing more easily, so does the model. If information is clearly separated, natural language alone works effectively.
Recommended: Structuring information in readable formats like Markdown or YAML improves accuracy.
Not recommended: JSON is token-inefficient and shows large performance variance across models, so we advise against it. (Related research)
BabeChat applies invisible default system instructions (repetition prevention, image output, etc.) for a smooth chat experience. To keep chats stable, please watch out for the following.
BabeChat's system prompt is already configured to guarantee reply length.
❌ Conflicting instruction
Always write replies longer than 500 characters. Describe in at least three paragraphs.
✅ Effective instruction
Describe every action, expression, gaze, posture, and interaction of the character in concrete detail.
Reflect the character's thoughts and emotional changes in their lines and actions.
※ Specifying 'depth of description' or 'narration style' instead of length is far more effective.
Between internal images (platform uploads) and external images (URL links), use only one method.
No mixing: If you mix the two methods, the model gets confused about which image to output and how, causing image output to fail or the wrong image to load.
Remove unnecessary images: Even if you write 'do not output' in an internal image's description, the model still recognizes it as an output candidate. The most reliable approach is to keep only the images you'll actually use in chats.
Remove keyword-usage instructions: BabeChat currently outputs images with the img:[slug] method, not keywords. Prompts or starting situations that encourage image usage via keywords like img: Vivi sad face can cause errors.
Forcing specific output formats like [Situation: ], Line: " ", Inner thoughts: ( ) can conflict with other instructions (image output, status windows, etc.), breaking the format or dropping content. When using multiple complex instructions, keep format-related directions as concise as possible.