diff --git a/.jules/bolt.md b/.jules/bolt.md
index a85f262..a6a99bb 100644
--- a/.jules/bolt.md
+++ b/.jules/bolt.md
@@ -1,3 +1,7 @@
 ## 2024-05-21 - [Blender Modal Redraws]
 **Learning:** Blender `modal` operators run frequently (e.g. every 0.3s). Calling `area.tag_redraw()` unconditionally in the loop forces constant viewport rendering, even when no state has changed, causing unnecessary CPU/GPU load.
 **Action:** Only call `tag_redraw()` when the operator actually processes data or updates the UI state (e.g., via a `refresh_needed` flag).
+
+## 2024-05-22 - [Google GenAI Client Reuse]
+**Learning:** Creating a new `google.genai.Client()` instance takes ~80ms. The original pipeline created 5 instances sequentially, adding ~400ms overhead. Reusing a single client instance saves this overhead and likely benefits from connection pooling.
+**Action:** Instantiate API clients once at the start of a pipeline and pass them down to utility functions using dependency injection.
diff --git a/operators.py b/operators.py
index e202f95..f0de249 100644
--- a/operators.py
+++ b/operators.py
@@ -90,10 +90,13 @@ def execute(self, context):
 
     def _run_pipeline(self, gemini_key, meshy_key, prompt, q):
         try:
+            # ⚡ Bolt: Reuse client to avoid overhead (saving ~400ms)
+            client = utils.get_client(gemini_key)
+
             q.put(("INFO", "Refining prompt...", ""))
 
             # Step 1: Refine prompt
-            refined = utils.refine_prompt(gemini_key, prompt)
+            refined = utils.refine_prompt(gemini_key, prompt, client=client)
             q.put(("REFINED", refined, ""))
             q.put(("INFO", "Prompt refined", ""))
 
@@ -110,7 +113,9 @@ def generate_view(view_name, view_prompt, input_ref=None):
 
                 # If it's the front view call, input_ref is None usually,
                 # but valid for subsequent calls
-                res_path = utils.generate_image(gemini_key, view_prompt, out, input_ref)
+                res_path = utils.generate_image(
+                    gemini_key, view_prompt, out, input_ref, client=client
+                )
                 q.put(("IMAGE", f"{view_name} done", res_path))
                 return res_path
 
diff --git a/utils.py b/utils.py
index b12dfd5..e3789f7 100644
--- a/utils.py
+++ b/utils.py
@@ -16,9 +16,10 @@ def get_client(api_key):
     return genai.Client(api_key=api_key)
 
 
-def refine_prompt(api_key, prompt):
+def refine_prompt(api_key, prompt, client=None):
     """Use Gemini to refine a prompt for 3D model generation."""
-    client = get_client(api_key)
+    if client is None:
+        client = get_client(api_key)
 
     instruction = (
         f"Refine this prompt for generating a high-quality 3D model reference image. "
@@ -36,7 +37,7 @@ def refine_prompt(api_key, prompt):
     return response.text.strip()
 
 
-def generate_image(api_key, prompt, output_path, input_image_path=None):
+def generate_image(api_key, prompt, output_path, input_image_path=None, client=None):
     """
     Generate an image using Gemini.
     If input_image_path is provided, use it as reference for the generation.
@@ -44,7 +45,8 @@ def generate_image(api_key, prompt, output_path, input_image_path=None):
     from google.genai import types
     from PIL import Image
 
-    client = get_client(api_key)
+    if client is None:
+        client = get_client(api_key)
 
     config = types.GenerateContentConfig(
         response_modalities=["Image"],
@@ -117,7 +119,8 @@ def generate_3d_meshy(api_key, image_paths):
         task_id = resp.json()["result"]
 
         # Adaptive polling: Check frequently at first (2s), then back off to 5s
-        # This reduces waiting time for fast jobs without spamming the API for slow ones.
+        # This reduces waiting time for fast jobs without spamming the API
+        # for slow ones.
         intervals = [2, 2, 2, 5]
         default_interval = 5