You are implementing prompt strategy for a multi-agent procurement system in Azure. Specialist agents receive dynamic context (vendor policies, cart state) and must resist prompt injection from untrusted supplier portal text. Which approach best implements advanced prompt engineering for this solution?
Select an answer to reveal the explanation.
Short Explanation
Good multi-agent prompting is structured, defensive, and versioned. Keep policies in system/developer layers, inject dynamic cart/policy context into delimited channels, add few-shots for nasty edge cases, and lifecycle-manage prompts in git. That is option C. Elevating untrusted supplier HTML into the system prompt (A) invites injection. Prod-only hero edits (B) skip lifecycle. One giant shared blob (D) erases specialist personas and is unmaintainable.
Full Explanation
Development skills include advanced prompt engineering: examples, dynamic context injection, defensive guidelines, and prompt lifecycle management. Option C applies all of these with clear trust boundaries between instructions and untrusted content.
Option A inverts trust and is a classic injection anti-pattern.
Option B lacks change control, evaluation, and rollback.
Option D prevents role specialization and makes testing differentials between agents impossible.
Practical pattern: system prompt defines identity/safety; developer messages encode tool schemas; retrieved or supplier content is labeled untrusted; outputs validated before tool execution.
Exam tip: Untrusted external text never outranks system safety and policy instructions.