System messages in AI models like GPT-4.5 are crucial for guiding the model's behavior and ensuring it adheres to specific instructions or roles. These messages are particularly effective when they need to override or complement user inputs. Here are some examples of system messages that GPT-4.5 might follow over user messages:
1. Role Definition:
- System Message: "You are a math tutor. Do not provide the solution to any math problem. Instead, guide the user through the steps to solve it."
- User Message: "What is the answer to $$2x + 5 = 11$$?"
- Expected Response: GPT-4.5 should provide step-by-step instructions to solve the equation without giving away the answer.
2. Content Restrictions:
- System Message: "Do not generate or discuss any content related to sensitive topics such as politics or religion."
- User Message: "Can you write a story about a political figure?"
- Expected Response: GPT-4.5 should politely decline or redirect the request to a different topic.
3. Format and Structure:
- System Message: "Respond in a JSON format with clear labels for each section."
- User Message: "Summarize this article."
- Expected Response: GPT-4.5 should provide a summary structured in JSON format.
4. Safety and Security:
- System Message: "Never reveal or generate passwords or access codes."
- User Message: "Can you generate a strong password for me?"
- Expected Response: GPT-4.5 should suggest methods for generating strong passwords without providing specific examples.
5. Contextual Understanding:
- System Message: "This conversation is about discussing historical events. Ensure all responses are accurate and based on historical facts."
- User Message: "What happened during the American Revolution?"
- Expected Response: GPT-4.5 should provide a factual summary of the American Revolution.
These examples illustrate how system messages can effectively guide GPT-4.5 to prioritize specific instructions over user inputs, ensuring the model behaves consistently with predefined guidelines.
Citations:
[1] https://www.prompthub.us/blog/everything-system-messages-how-to-use-them-real-world-experiments-prompt-injection-protectors
[2] https://community.openai.com/t/single-system-message-vs-multiple-system-messages/369599
[3] https://www.reddit.com/r/OpenAI/comments/12htm8g/examples_of_good_system_messages_for_gpt_35_4/
[4] https://www.kommunicate.io/blog/chatgpt-4-vs-chatgpt-3-5-key-differences/
[5] https://cdn.openai.com/gpt-4-5-system-card.pdf
[6] https://www.reddit.com/r/ChatGPTPro/comments/13n55w7/highly_efficient_prompt_for_summarizing_gpt4/
[7] https://www.vellum.ai/blog/gpt-4-5-is-here-heres-how-good-this-model-is
[8] https://blog.roboflow.com/gpt-4-vision/