System Prompts Explained Quiz

Answers and explanations appear as you take the quiz.

1. What is a system prompt?
   • The first message a user types in a chat
   • Instructions supplied in a privileged message that set the model's role, rules and context for the conversation
   • The model's training dataset
   • An error message shown when a model fails
2. Why do models usually give system instructions more weight than conflicting user requests?
   • A hard-coded rule in the API server deletes conflicting user messages
   • System prompts use a different programming language
   • System prompts are processed by a separate model
   • The models were trained on examples where system instructions take precedence
3. How does Anthropic's Messages API accept a system prompt?
   • As a top-level system field
   • As a user message starting with "SYSTEM:"
   • As an uploaded file
   • Only through a web dashboard
4. Why is the system prompt sent with every API request?
   • To verify the user's identity
   • Because the model is stateless between calls and does not remember earlier requests
   • Because the law requires it
   • To train the model on each request
5. What does a chat template such as ChatML do?
   • Encrypts the conversation
   • Translates prompts into English
   • Converts the roles in a conversation into one token sequence with special tokens marking system, user and assistant segments
   • Limits responses to 100 words
6. Which statement about system prompt security is accurate?
   • System prompts cannot be extracted under any circumstances
   • Putting API keys in the system prompt is the safest place to store them
   • Prompt injection and extraction attacks can succeed, so real security has to be enforced outside the model
   • User messages are never able to influence system-level behavior
7. Why should static instructions come before variable content in a system prompt?
   • It lets prompt caching reuse the stable prefix, which reduces cost and latency
   • Models ignore text that comes after variable content
   • APIs reject prompts that do otherwise
   • It makes the model answer in a friendlier tone
8. In the chain of command described in OpenAI's Model Spec, how should instructions found inside tool outputs or retrieved documents be handled?
   • They outrank platform rules
   • They are equal to developer instructions
   • They rank above user instructions but below the platform
   • They carry no authority by default and should be treated as information