diff --git a/plugins/hookify/core/config_loader.py b/plugins/hookify/core/config_loader.py
index fa2fc3e36..73d1bbcfd 100644
--- a/plugins/hookify/core/config_loader.py
+++ b/plugins/hookify/core/config_loader.py
@@ -40,6 +40,9 @@ class Rule:
     action: str = "warn"  # "warn" or "block" (future)
     tool_matcher: Optional[str] = None  # Override tool matching
     message: str = ""  # Message body from markdown
+    silent: bool = False  # If true, suppress systemMessage (banner) and
+                          # deliver the warning only via additionalContext
+                          # (assistant-facing self-correction mode).
 
     @classmethod
     def from_dict(cls, frontmatter: Dict[str, Any], message: str) -> 'Rule':
@@ -80,7 +83,8 @@ def from_dict(cls, frontmatter: Dict[str, Any], message: str) -> 'Rule':
             conditions=conditions,
             action=frontmatter.get('action', 'warn'),
             tool_matcher=frontmatter.get('tool_matcher'),
-            message=message.strip()
+            message=message.strip(),
+            silent=bool(frontmatter.get('silent', False)),
         )
 
 
diff --git a/plugins/hookify/core/rule_engine.py b/plugins/hookify/core/rule_engine.py
index 51561c39e..e4e8c58e4 100644
--- a/plugins/hookify/core/rule_engine.py
+++ b/plugins/hookify/core/rule_engine.py
@@ -73,7 +73,8 @@ def evaluate_rules(self, rules: List[Rule], input_data: Dict[str, Any]) -> Dict[
                 return {
                     "hookSpecificOutput": {
                         "hookEventName": hook_event,
-                        "permissionDecision": "deny"
+                        "permissionDecision": "deny",
+                        "permissionDecisionReason": combined_message,
                     },
                     "systemMessage": combined_message
                 }
@@ -85,10 +86,25 @@ def evaluate_rules(self, rules: List[Rule], input_data: Dict[str, Any]) -> Dict[
 
         # If only warnings, show them but allow operation
         if warning_rules:
-            messages = [f"**[{r.name}]**\n{r.message}" for r in warning_rules]
-            return {
-                "systemMessage": "\n\n".join(messages)
-            }
+            all_messages = [f"**[{r.name}]**\n{r.message}" for r in warning_rules]
+            visible_messages = [f"**[{r.name}]**\n{r.message}"
+                                for r in warning_rules
+                                if not getattr(r, 'silent', False)]
+            response = {}
+            # Silent rules suppress the user-facing banner; non-silent rules
+            # still show systemMessage. If ALL matching rules are silent, no
+            # banner is emitted at all.
+            if visible_messages:
+                response["systemMessage"] = "\n\n".join(visible_messages)
+            # PreToolUse additionalContext supported since Claude Code v2.1.9.
+            # PostToolUse has supported it since earlier. Inject so Claude
+            # (not just the user console) sees the educational message.
+            if hook_event in ('PreToolUse', 'PostToolUse'):
+                response["hookSpecificOutput"] = {
+                    "hookEventName": hook_event,
+                    "additionalContext": "\n\n".join(all_messages),
+                }
+            return response
 
         # No matches - allow operation
         return {}