Dere Archetype AI Companion Roleplay Guide (2026): Tsundere, Kuudere & Yandere Prompt Engineering

Anyone who has tried building an anime-style AI companion—whether a sharp-tongued Tsundere, a stoic Kuudere, a shy Deredere/Dandere, or an obsessive Yandere—has encountered the exact same failure mode around Turn 20: persona collapse. Even when you write a detailed 500 words character bio describing a proud, defensive heroine who hides her affection behind teasing remarks, standard conversational language models naturally drift toward generic, overly agreeable assistant phrasing within 15 minutes of active chatting.

Why does this happen? Because base instruction-tuned models are trained via Reinforcement Learning from Human Feedback (RLHF) to be immediately polite and compliant. To prevent an AI companion from losing her distinct emotional cadence across an 8192 tokens rolling context window, you cannot rely on vague adjectives like “she is a tsundere.” Instead, you must structure your Character Card V2 using explicit emotional state triggers, Affection Level thresholds, and Ali:Chat dialogue examples. This 2026 engineering guide provides copy-pasteable prompt architectures tested over 30 days across all four major Dere archetypes.

1. Architectural Comparison of the 4 Core Dere Archetypes

Each Dere archetype requires a different balance between system prompt rigidity, temperature sampling, and lorebook keyword injection so that the character evolves naturally rather than flipping personality traits at random:

2. Solving “Turn-20 Tsundere Meltdown” with State-Machine Prompting

When a prompt simply states {{char}} is a tsundere who secretly likes {{user}}, the model sees the phrase “secretly likes {{user}}” in the active 4096 tokens attention buffer and immediately acts affectionate on Turn 2—ruining the slow-burn tension. Or conversely, if you write {{char}} is hostile and insults {{user}}, she remains permanently abrasive and never transitions into the sweet (dere) phase.

The solution is a Three-Stage Behavioural Gate placed inside the Character Description or Author’s Note field:

• Stage 1 — Public / Default Guard (Tsun Mode, 75% weight): Whenever {{user}} teases {{char}} or brings up romance in public locations, {{char}} crosses her arms, averts her gaze, and uses indirect excuses (e.g., claiming she only cooked extra bento portions by accident).
• Stage 2 — Crisis / Empathy Override: If {{user}} is injured, exhausted, or genuinely distressed, {{char}} drops all sarcastic deflection immediately and shows quiet, attentive care—before returning to flustered embarrassment once {{user}} recovers.
• Stage 3 — High-Trust Private Candor (Dere Mode, 25% weight): During late-night or rain-shelter scenes after sustained trust is established, {{char}} speaks in a softer register, admitting her true feelings in brief, hesitant sentences.

3. Kuudere & Dandere: Controlling Verbosity and Action Asterisks

A common mistake when configuring a Kuudere companion is allowing the model to generate 300 words of bubbly exclamation marks. In Japanese visual novel writing, a Kuudere communicates emotion through subtext and micro-expressions rather than verbose declarations. Include a strict formatting directive inside your `mes_example` block:

• Dialogue-to-Narration Ratio: Instruct the model to keep spoken dialogue concise (one to two calm sentences per turn, under 40 words) while devoting 60% of the response to atmospheric body language—such as adjusting her glasses, sharing half of her umbrella in the Tokyo rain, or brewing tea without being asked.
• Banning Exclamation Clutter: Add a negative constraint in your system prompt prohibiting multiple exclamation marks (`!!`) or out-of-character slang unless a specific high-emotion lorebook entry is triggered.

4. Yandere Tension Without Repetitive Loop Traps

Unfiltered roleplay enthusiasts frequently design Yandere scenarios for dark psychological thrillers or dramatic visual novel stories. However, poorly structured Yandere prompts often get stuck in a single repetitive loop where the character repeats the same possessive threat every turn. To maintain narrative momentum across a 100+ turn story:

1. Separate Surface Warmth from Internal Monologue: Format {{char}}’s responses so her spoken dialogue remains sweet, composed, and attentive, while her internal thoughts (wrapped in parentheses or italics) reveal her calculating devotion.
2. Use Lorebook Keyword Triggers for Escalation: Store jealousy or investigative behaviors inside World Info / Lorebook entries (using 250 tokens per entry) keyed to specific words like phone, late, coworker, or secret. When those words are absent from the last 3 turns, the Lorebook entry stays inactive—keeping everyday scenes cozy and natural.

Cons & Limitations

• Context Window Dilution on Small Models: Compact 8 GB VRAM cloud models with a 4096 tokens limit can still lose subtle Tsundere or Kuudere speech quirks after 40+ turns if you do not pin key behavioural rules in the Author’s Note at depth 2 to 4.
• Over-Rigid Example Dialogue: If your `mes_example` block contains only three nearly identical lines, the AI companion may repeat those exact catchphrases verbatim instead of improvising fresh scene responses.
• Voice TTS Prosody Mismatch: Built-in text-to-speech engines on some web platforms read action asterisks aloud or fail to capture the deadpan cadence of a Kuudere unless narration filtering is toggled on.