AIF-C01 Practice Question 97: System Prompt Persona and Scope
AWS Certified AI Practitioner (AIF-C01) · Test 10 · question 7 of 10
A bank is building a support assistant on a foundation model. It must always reply in a formal tone as the bank's assistant and must politely decline topics that are not about banking. These rules must apply on every turn of a conversation and take precedence over what users write. Where should the developer put these instructions?
Show answer and explanation
Answer:
The system prompt provides *"instructions to the model that define how it will respond to end users of your application."* It can set a persona, define allowable and unallowable content and specify output style. It *"will supersede other instructions provided in individual user prompts, and will carry over across all user turns"* (option A). AWS also suggests instructing the model to politely decline requests that fall outside its scope.
- B is wrong because manually appending rules to each user message is fragile, easy to forget, and does not give the rules precedence over what the user writes. - C is wrong because fine-tuning changes model behavior through training. It is slow and costly for a rule that can be stated in text, and it does not give the rules precedence over user input. - D is wrong because a few-shot example only demonstrates one pattern. It is not an instruction with precedence, and it is easily outweighed by what users write.
References: Using the system role; ConverseStream (the system field)