diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index b436f4dc6d..44c700bf3d 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -45,8 +45,8 @@ _INTENT_SIGNAL = re.compile(
     r"\b(?:now i|next i)\b"
     r")"
 )
-_MAX_REPROMPTS = 1
-_REPROMPT_MAX_CHARS = 500
+_MAX_REPROMPTS = 3
+_REPROMPT_MAX_CHARS = 2000
 
 # ── Pre-compiled patterns for GGUF shard detection ───────────
 _SHARD_FULL_RE = re.compile(r"^(.*)-(\d{5})-of-(\d{5})\.gguf$")
@@ -2659,8 +2659,9 @@ class LlamaCppBackend:
                                 {
                                     "role": "user",
                                     "content": (
-                                        "Please use the available tools to complete "
-                                        "the task instead of describing what to do."
+                                        "STOP. Do NOT write code or explain. "
+                                        "You MUST call a tool NOW. "
+                                        "Call web_search or python immediately."
                                     ),
                                 }
                             )
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 8f058bd0f2..ced24c1d5f 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -96,8 +96,10 @@ router = APIRouter()
 
 # Appended to tool-use nudge to discourage plan-without-action
 _TOOL_ACTION_NUDGE = (
-    " Always call tools directly."
+    " IMPORTANT: Always call tools directly -- never write code yourself."
     " Never describe what you plan to do -- just call the tool immediately."
+    " For any code request, call the python tool. For any factual question, call web_search."
+    " Do NOT output code blocks -- use the python tool instead."
 )
 
 # Regex for stripping leaked tool-call XML from assistant messages/stream