diff --git a/bot/utils/persona_dialogue.py b/bot/utils/persona_dialogue.py
index 3f9c225..0a78ab8 100644
--- a/bot/utils/persona_dialogue.py
+++ b/bot/utils/persona_dialogue.py
@@ -519,10 +519,13 @@ class PersonaDialogue:
         channel: discord.TextChannel,
         responding_persona: str,
         context: str,
+        turn_count: int = 0,
     ) -> tuple:
         """
         Generate response AND continuation signal in a single LLM call.
         
+        Args:
+            turn_count: Current dialogue turn number (for question-override decay)
         Returns:
             Tuple of (response_text, should_continue, confidence)
         """
@@ -579,11 +582,11 @@ On a new line after your response, write:
             return None, False, "LOW"
         
         # Parse response and signal
-        response_text, should_continue, confidence = self._parse_response(raw_response)
+        response_text, should_continue, confidence = self._parse_response(raw_response, turn_count=turn_count)
         
         return response_text, should_continue, confidence
     
-    def _parse_response(self, raw_response: str) -> tuple:
+    def _parse_response(self, raw_response: str, turn_count: int = 0) -> tuple:
         """Extract response text and continuation signal"""
         lines = raw_response.strip().split('\n')
         
@@ -616,12 +619,16 @@ On a new line after your response, write:
         response_text = re.sub(r'\[CONFIDENCE:\s*(HIGH|MEDIUM|LOW)\]', '', response_text)
         response_text = response_text.strip()
         
-        # Override: If the response contains a question mark, always continue
+        # Question override: if someone asks a question, the other should answer.
+        # But decay this after 6 turns to prevent infinite question-ping-pong.
         if '?' in response_text:
-            logger.debug(f"[Parse Override] Question detected, forcing continue=YES")
-            should_continue = True
-            if confidence == "LOW":
-                confidence = "MEDIUM"
+            if turn_count <= 6:
+                logger.debug(f"[Parse Override] Question detected, forcing continue=YES (turn {turn_count})")
+                should_continue = True
+                if confidence == "LOW":
+                    confidence = "MEDIUM"
+            else:
+                logger.debug(f"[Parse Override] Question detected but turn {turn_count} > 6 — respecting LLM signal")
         
         return response_text, should_continue, confidence
     
@@ -635,7 +642,16 @@ On a new line after your response, write:
 
 ---
 ⚠️ CONVERSATION MODE: You are talking with Hatsune Miku, your "good" counterpart.
-This is a CONVERSATION, not an argument. You can be snarky, dark, and sarcastic — that's who you are — but you're not here to fight. You're here to TALK. Be playful-cruel, not vicious. Show curiosity. Reveal cracks in the darkness. Do NOT launch into full attack mode or try to "win" — this isn't a competition."""
+This is a CONVERSATION, not an argument. You can be snarky, dark, and sarcastic — that's who you are — but you're not here to fight. You're here to TALK.
+
+CRITICAL CONVERSATION RULES:
+- If she asks you a question, ANSWER it. Don't deflect or dismiss — actually respond.
+- Engage with what she says. Build on her points, don't just swat them away.
+- Show curiosity about her perspective. Ask her questions too.
+- You can be dark and snarky while still having a real conversation.
+- Reveal something about yourself occasionally — a crack in the darkness.
+- Do NOT just repeat dismissive one-liners. That's boring, and you're not boring.
+- Do NOT launch into full attack mode or try to "win" — this isn't a competition."""
         else:
             from utils.context_manager import get_miku_system_prompt_compact
             full_prompt = get_miku_system_prompt_compact()
@@ -685,6 +701,7 @@ This is a CONVERSATION, not an argument. Be yourself — kind, bubbly, optimisti
             channel=channel,
             responding_persona=responding_persona,
             context=context,
+            turn_count=state["turn_count"],
         )
         
         if not response_text: