diff --git a/wolverine.py b/wolverine.py
index f1d910d..d9eff4a 100644
--- a/wolverine.py
+++ b/wolverine.py
@@ -1,18 +1,27 @@
 import difflib
-import fire
 import json
 import os
 import shutil
 import subprocess
 import sys
-import openai
+
+import fire
+from dotenv import load_dotenv
 from termcolor import cprint
 from dotenv import load_dotenv
 
 
-# Set up the OpenAI API
 load_dotenv()
-openai.api_key = os.getenv("OPENAI_API_KEY")
+
+import openai
+
+openai.api_key = os.environ.get("OPENAI_API_KEY")
+
+DEFAULT_MODEL = os.environ.get("DEFAULT_MODEL", "gpt-4")
+
+
+with open("prompt.txt") as f:
+    SYSTEM_PROMPT = f.read()
 
 
 def run_script(script_name, script_args):
@@ -26,7 +35,47 @@ def run_script(script_name, script_args):
     return result.decode("utf-8"), 0
 
 
-def send_error_to_gpt(file_path, args, error_message, model):
+def send_error_to_gpt(file_path, args, error_message, model=DEFAULT_MODEL):
+    def json_validated_response(model, messages):
+        """
+        This function is needed because the API can return a non-json response.
+        This will run recursively until a valid json response is returned.
+        """
+        response = openai.ChatCompletion.create(
+            model=model,
+            messages=messages,
+            temperature=0.5,
+        )
+        messages.append(response.choices[0].message)
+        content = response.choices[0].message.content
+        # see if json can be parsed
+        try:
+            json_start_index = content.index(
+                "["
+            )  # find the starting position of the JSON data
+            json_data = content[
+                json_start_index:
+            ]  # extract the JSON data from the response string
+            json_response = json.loads(json_data)
+        except (json.decoder.JSONDecodeError, ValueError) as e:
+            cprint(f"{e}. Re-running the query.", "red")
+            # debug
+            cprint(f"\n\GPT RESPONSE:\n\n{content}\n\n", "yellow")
+            # append a user message that says the json is invalid
+            messages.append(
+                {
+                    "role": "user",
+                    "content": "Your response could not be parsed by json.loads. Please restate your last message as pure JSON.",
+                }
+            )
+            # rerun the api call
+            return json_validated_response(model, messages)
+        except Exception as e:
+            cprint(f"Unknown error: {e}", "red")
+            cprint(f"\n\GPT RESPONSE:\n\n{content}\n\n", "yellow")
+            raise e
+        return json_response
+
     with open(file_path, "r") as f:
         file_lines = f.readlines()
 
@@ -35,12 +84,7 @@ def send_error_to_gpt(file_path, args, error_message, model):
         file_with_lines.append(str(i + 1) + ": " + line)
     file_with_lines = "".join(file_with_lines)
 
-    with open("prompt.txt") as f:
-        initial_prompt_text = f.read()
-
     prompt = (
-        initial_prompt_text +
-        "\n\n"
         "Here is the script that needs fixing:\n\n"
         f"{file_with_lines}\n\n"
         "Here are the arguments it was provided:\n\n"
@@ -52,26 +96,27 @@ def send_error_to_gpt(file_path, args, error_message, model):
     )
 
     # print(prompt)
-    response = openai.ChatCompletion.create(
-        model=model,
-        messages=[
-            {
-                "role": "user",
-                "content": prompt,
-            }
-        ],
-        temperature=1.0,
-    )
+    messages = [
+        {
+            "role": "system",
+            "content": SYSTEM_PROMPT,
+        },
+        {
+            "role": "user",
+            "content": prompt,
+        },
+    ]
 
-    return response.choices[0].message.content.strip()
+    return json_validated_response(model, messages)
 
 
-def apply_changes(file_path, changes_json):
+def apply_changes(file_path, changes: list):
+    """
+    Pass changes as loaded json (list of dicts)
+    """
     with open(file_path, "r") as f:
         original_file_lines = f.readlines()
 
-    changes = json.loads(changes_json)
-
     # Filter out explanation elements
     operation_changes = [change for change in changes if "operation" in change]
     explanations = [
@@ -114,7 +159,7 @@ def apply_changes(file_path, changes_json):
             print(line, end="")
 
 
-def main(script_name, *script_args, revert=False, model="gpt-4"):
+def main(script_name, *script_args, revert=False, model=DEFAULT_MODEL):
     if revert:
         backup_file = script_name + ".bak"
         if os.path.exists(backup_file):
@@ -140,11 +185,11 @@ def main(script_name, *script_args, revert=False, model="gpt-4"):
             print("Output:", output)
 
             json_response = send_error_to_gpt(
-                    file_path=script_name,
-                    args=script_args,
-                    error_message=output,
-                    model=model,
+                file_path=script_name,
+                args=script_args,
+                error_message=output,
             )
+
             apply_changes(script_name, json_response)
             cprint("Changes applied. Rerunning...", "blue")