diff --git a/environments/benchmarks/terminalbench_2/default.yaml b/environments/benchmarks/terminalbench_2/default.yaml
index 2a3647f908b..9797eb0918c 100644
--- a/environments/benchmarks/terminalbench_2/default.yaml
+++ b/environments/benchmarks/terminalbench_2/default.yaml
@@ -14,19 +14,6 @@
 #       --openai.model_name anthropic/claude-sonnet-4
 
 env:
-  system_prompt: |
-    You are an expert software engineer and Linux systems administrator solving a task in an isolated Linux environment. You have terminal and file tools available.
-
-    Before each action, briefly analyze what you see and plan your next step. After running a command, read its output carefully before proceeding.
-
-    Approach:
-    - Start by exploring: read the task files, check the environment, understand what's given.
-    - Break complex problems into steps. Solve and verify each step before moving on.
-    - After making changes, always test them — run the test command, check the output.
-    - If something fails, read the error, diagnose the cause, and try a different approach. Do not give up or repeat the same failing command.
-    - Do not stop until you have verified your solution works.
-
-    When to stop: Once you believe your solution is complete and you have verified it works (e.g. the program runs correctly, the output looks right, the file is in place), respond with a plain text message summarizing what you did. Do NOT make any more tool calls after that.
   enabled_toolsets: ["terminal", "file"]
   max_agent_turns: 100
   max_token_length: 32000
diff --git a/environments/benchmarks/terminalbench_2/terminalbench2_env.py b/environments/benchmarks/terminalbench_2/terminalbench2_env.py
index 748d706e249..6e22ddaaf93 100644
--- a/environments/benchmarks/terminalbench_2/terminalbench2_env.py
+++ b/environments/benchmarks/terminalbench_2/terminalbench2_env.py
@@ -56,6 +56,7 @@ from atroposlib.envs.base import EvalHandlingEnum
 from atroposlib.envs.server_handling.server_manager import APIServerConfig
 from pydantic import Field
 
+from agent.prompt_builder import DEFAULT_AGENT_IDENTITY
 from environments.agent_loop import AgentResult, HermesAgentLoop
 from environments.hermes_base_env import HermesAgentBaseEnv, HermesAgentEnvConfig
 from environments.tool_context import ToolContext
@@ -146,11 +147,14 @@ MODAL_INCOMPATIBLE_TASKS = {
 # Injected as a user message when the model responds with plain text instead of
 # calling a tool or including a <task_status> tag.
 _FORMAT_NUDGE_MESSAGE = (
-    "Your response must be one of the following:\n"
-    "1. A tool call (e.g. terminal, read_file, write_file) to continue working on the task.\n"
-    "2. <task_status>DONE</task_status> — if you have fully completed the task.\n"
-    "3. <task_status>UNFINISHED</task_status> — if you are unable to complete the task.\n\n"
-    "Plain text responses are not accepted. Please continue working or report your final status."
+    "You wrote a plain text response instead of using your tools. "
+    "Plain text responses do not affect the environment — nothing was executed or saved.\n\n"
+    "You MUST use your tools (terminal, read_file, write_file) to actually complete the task. "
+    "Do not describe what you would do — execute it now by making tool calls.\n\n"
+    "If you have already completed all required work using tools in previous turns, "
+    "respond with exactly: <task_status>DONE</task_status>\n"
+    "If you have exhausted all approaches and cannot make further progress, "
+    "respond with exactly: <task_status>UNFINISHED</task_status>"
 )
 
 # Maximum number of format nudges before giving up and moving on to scoring.
@@ -279,7 +283,7 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
             max_agent_turns=60,
             max_token_length=16000,
             agent_temperature=0.6,
-            system_prompt=None,
+            system_prompt=DEFAULT_AGENT_IDENTITY,
             # Modal backend for per-task cloud-isolated sandboxes
             terminal_backend="modal",
             terminal_timeout=300,  # 5 min per command (builds, pip install, etc.)