Which of the following is a direct effect of system messages on prompt design?

• A. They alter the model's underlying architecture.
• B. They set the persona, tone, and safety constraints guiding outputs.
• C. They automatically correct user input.
• D. They influence only the initial token embedding.

Answer: (B)

System messages establish the guardrails that shape how the model responds. They directly set the persona, tone, and safety constraints guiding outputs, so the model replies in a consistent character and adheres to safety and content boundaries across the entire interaction. This is why it’s the best fit: it describes the purposeful boundaries you’re encoding to steer every answer.

They don’t change the model’s underlying architecture—that’s fixed by training and cannot be altered by prompts. They also don’t automatically fix user input; while a system message can instruct the model to do so, automatic correction is not a direct result of the system prompt itself. And the influence isn’t limited to the very first tokens—the system message helps guide every response throughout the conversation, not just the initial embedding.