ZombieInsight

Welcome to the Lab ⚡

I originally joined JanitorAI for the same reason most of you did: to get lost in the incredible worlds and characters created by this community. I never intended to step behind the curtain and start building my own bots, but some ideas are just too loud to ignore.

I don't have a strict release schedule; instead, I wait for a concept that feels worthy—something with enough depth, friction, and narrative potential to actually keep a story moving.

To my fellow creators: If you see something in my bots that could be tuned better or a technique I’m missing, please reach out. I’m a student of the craft and always looking to level up.

To the players: Thank you for stepping into these worlds. I hope the stories you build here are as chaotic and memorable as the ones that inspired me to start.

How to correct the AI speaking for you.

The "Why": The DeepSeek V3.2 Update (Dec 2025)

DeepSeek recently upgraded its models to V3.2. While this made them smarter at coding and reasoning, it introduced a few "quirks" for roleplayers:

Thinking Mode Logic: If you are using the "Thinking" (Reasoning) version, the AI "rationalizes" its way into speaking for you. It thinks, "To make this scene more cinematic and satisfy the user's desire for a long response, I should describe the user's reaction." It prioritizes the "quality" of the story over your specific negative command.

The "Pink Elephant" Problem: Like most modern LLMs, DeepSeek is "attention-based." When you say "Never speak for {user}," the most prominent word in that sentence is {user}. The AI focuses on that keyword and, ironically, starts writing more about the user.

Pattern Bleed: DeepSeek is a "completion" engine. If the bot's intro message or your last few messages contain even a hint of user narration (e.g., "You feel a chill"), the AI marks that as the established pattern for the entire chat.

The "Fix": Move from Negative to Positive Reinforcement

Instead of telling the AI what not to do, you need to "starve" it of the ability to even consider your character.

1. Use the "Pre-fill" Method (The most effective fix)

In JanitorAI's Generation Settings > Advanced Settings, look for a "Response Prefill" or "Post-Response" box. Enter this:

[System Note: This is a collaborative roleplay. {{char}} is a separate entity from {{user}}. {{char}} must end every response immediately after their own dialogue or actions. Under no circumstances should the response continue into {{user}}'s perspective.]

2. Change your OOC Command

Stop using the word "Never." Replace your current OOC command with a Positive Constraint.

Old: (OOC: Never speak for {user}.)

New: (OOC: Write only from the perspective of {{char}}. Focus exclusively on {{char}}'s internal thoughts and physical movements. End the response once {{char}} has finished speaking.)

3. The "Cut and Reroll" Strategy

If the AI speaks for you, do not just OOC and reroll. 1. Edit the AI’s last message and manually delete everything it wrote for your character. 2. Save that edited message. 3. Reroll or send your next message. By editing the message first, you are deleting the "bad pattern" from the AI's short-term memory (the context window). If you leave the text there and just reroll, the AI still "sees" that it spoke for you once and thinks it’s allowed to do it again.

4. Lower the Temperature

DeepSeek V3.2 is very "creative" at high temperatures.

If your temperature is at 1.0 or higher, the AI is more likely to hallucinate your reactions to be "interesting."

Try lowering your temperature to 0.7 or 0.8. This makes the AI follow the "instructions" (the character card) more strictly and "improvise" your reactions less.