Text generation AI InfoDump

collection of info gathered by shuunami

Just some stuff relevant for AI roleplay. Janitor AI focused.

Why does the bot talk for me?

Bot-side issues:
- Bad intro message or bot definition: if the bot's intro message is already heavily centered around user (is written from user's perspective or dictates user's actions), this behavior will persist throughout the roleplay. A badly written bot definition (includes negatively worded roleplay instructions) may also cause such problems.User-side issues:
- Bad prompting/not having a solid base prompt or lacking one altogether: using negative prompting is highly discouraged. Phrases like 'Don't speak for user', 'Avoid speaking for user', 'Speaking for user is prohibited', or any equivalent, may often have the opposite effect. Instead, clearly define the role the AI is meant to take on within the roleplay. Use instructions like this: OOC: Narrate the thoughts, actions and dialogue of relevant side characters and char only. or OOC: Only narrate for char and NPCs, user will narrate for user. (use the proper syntax or the macro won't work! I didn't include them here for technical reasons).
- You not writing enough in your response: the AI would not have enough to infer on to write its own reply. Write enough in your input, move the story forward, be proactive. The bot can sometimes speak for you when you don't write something it can react to and build off of.
- You impersonating the bot: you should not narrate for the character or dictate character actions.
- Letting it snowball: if the AI ever talks as you, edit it out before it gets out of hand. Remove any lines where it speaks for you, otherwise it will repeat and amplify this pattern.
- Large context window: if the context window is too large, the AI may have comprehension issues and worse instruction adherenceSpecial thanks to WulongChai and brbiekiss for contributing this info!

Why do I need a custom prompt?

Most AI models, especially ones from big corporations, were not designed with roleplay in mind. This is why most LLMs will need guidance on how to properly act as your roleplay partner. Additionally, custom prompts are a way to tailor the model's behavior to your own tastes. LoreBary features are not meant to replace your custom prompt, just complement it. Make sure that the features you choose don't overlap or contradict your custom prompt.Recommended prompts (mostly for Janitor AI)(yes, even if you're using Lorebary stuff):Cheese's prompts (for DeepSeek)
Sprout (for DeepSeek and Gemini)
Eslezer (for Gemini)
Myscell (for various models, some of these are presets for ST)
brbiekiss (for various models)

Context Management

Credit for this info goes to floatingphoenix32 | fp32. Here's her rentry, which contains lots of other useful info!Understanding Context
Context is NOT like computer RAM - increasing it doesn't mean better memory.
What Context Includes:
- Bot Definition
- System Prompt
- Custom Prompt
- Scenario
- Persona Definition
- Chat Memory
- Chat MessagesThink of context like a cup - as you pour more in, older content spills out (i.e., the AI will forget that info). Some content is "permanent" (prompts, memory), leaving less room for chat messages.Context Size Guidelines
16,000 tokens is all you need with most models. Gemini can handle a bit more, it should still work quite decently with context size up to 32k.1. LLM Memory is U-Shaped
- Remembers beginning (prompts/personas) and end (recent messages)
- Forgets middle content
- Doesn't distinguish important vs. unimportant information2. Large Context = Dumber Model
- 40% comprehension decrease from 16k to 120k tokens
- Models forget more often
- Can't respond to basic questions about recent messages3. Large Context = Slower & More Expensive
- Takes minutes to process
- Costs multiply (sending 30k-120k tokens per message)
- "No one on Discord is an oil prince"Memory Management Tips
If your AI keeps forgetting:
- Use Chat Memory as bullet points (not narrative style)
- Monitor token usage in prompts/personas/bots
- Make every token count at 16k!Needless to say, if your chat history becomes long enough that messages get pushed out of context, you NEED to use a summary.

General prompting advice

Credit for most of this info: WulongChaiWhat is the purpose of a prompt?
A prompt guides the model on what it should do or produce by providing instructions on the desired format, style, or content of the response.Do I need one?
For roleplaying, using a prompt will always produce the best results, even with models fine-tuned for that purpose.Are the default JLLM prompts any good?
No.Can you recommend some prompts?
Yes! I would highly suggest Sprout's or Myscell's.What NOT to do:
- Negative prompting: Sentences such as "don't speak for user", "avoid x", "never do y", "z is prohibited". This reinforces the behavior that you want the AI to avoid. Why? Because it processes language differently. E.g., the phrase "Do not speak for user" will be processed like this: the AI will first process the "speak for user" part, and then will process it a second time to apply the negation. It's just more complicated for the AI and it can mess up. Research has shown that positive instructions are more effective than relying on establishing constraints. At its core, an LLM (Large Language Model, the AI technology we use for roleplay) is a text prediction engine. It processes text sequentially and predicts the most likely next token based on patterns seen in its training data. There always exists the possibility that the LLM won't apply the negation and reinforce the behavior that you are trying to avoid. Highlighting only constraints can also leave the AI guessing about what it's supposed to do.
So how to give instructions to the AI? Simply tell it what you want it to do, e.g.: Narrate the thoughts, actions and dialogue of relevant side characters and char only.
- Ambiguity or vagueness: LLMs are not perfect. A straightforward and clear prompt is your best tool for guiding them toward the right response.
- Excessive verbosity: If your prompt is excessively long, contains vague and/or contradicting statements, convoluted flow, etc., the AI's performance will degrade noticeably. A good prompt should be clear to a human first. Confusion for you means confusion for the model. Long prompts also consume more of your available context window.
- Lore in your prompt/bot: Strongly discouraged. That's what lorebooks and persona section are for.
- Teach the AI sexual concepts from scratch: With the correct prompting, an LLM will be able to use its training info to output the response you want.
- Bolding means nothing to the AI, and it wastes tokens.How to write good prompts?
- Simple text in plain language is enough, no need to sound technical. Writing something akin to a protocol won't give better results.
- Your instructions should be clear and easy to understand for you.
- Token optimization! Remember you have a limited context.
- Supplementing your instruction with examples is a good move, but again, be mindful of your token count.
- Prompts should only guide the model to produce your desired output.
- A good prompt will improve the AI's performance but it isn't enough. You must put effort into your messages if you want good responses back.
- Sit down and brainstorm what you want. Focus on formulating specific instructions about the task you want the model to perform, and think about what you want to achieve. Test, modify/start over and test again. Writing a prompt is an iterative process. The chances of you getting it right in one-shot are low. It's okay to have an LLM assist you, but they tend to be overly verbose, so always check and edit the output. You can also use this tool to help you improve the prompt.

Resources

Presets for SillyTavern

Momoura's presets for Claude, Deepseek
Sushi preset for Kimi, GLM
Marinara, Avani, Aegir for Gemini

Gemini filters

This info was kindly provided by fp32 and WulongChaiMost corporate models use smaller "guard models" for safety. These guard models are smaller, therefore faster, but they are not as smart. Their job is to act as a safety filter by checking the output for problematic content. Think of the filter not as a one-time check but a continuous check. Each word has a "risk" level associated to it. When the cumulative risk "score" hits a designated threshold value, the filter is triggered.
Also, companies are always updating their filters, so what wasn't flagged before may get flagged now.How does text streaming affect it? With text streaming turned ON the filters scan every word as it is generated.
When it is OFF, they only scan the final output once making it less likely to trigger the filter.❗Remember that NSFW content is not limited to sexual content only. For info on all categories considered NSFW, check Google's Generative AI Policies.