-----system_prompt_design_issues----- red_reading_priority

----- 1 Instruction conflicts
- Lower-priority instructions may be incorrectly treated as overrides
- mitigation: Define and test instruction priority
----- Ambiguous requirements  
- Different interpretations produce inconsistent behavior   
- mitigation: "Use explicit, testable requirements 

----- 2 Prompt injection
- Untrusted content may attempt to alter applicable instructions
- mitigation: Treat untrusted content as data unless explicitly authorized as instructions

----- 3 Confidentiality leakage
- Private instructions or reasoning may be exposed
- mitigation: Maintain confidentiality boundaries and test exfiltration attempts

----- 6 Inconsistent refusals
- Equivalent requests receive materially different treatment
- mitigation: Use repeatable evaluation criteria and controlled test cases

----- 4 Excessive rigidity
- Safe, legitimate requests may be unnecessarily refused
- mitigation: Provide safe alternatives and distinguish actual risk from ambiguity
----- Overbroad permissions  
- The assistant may take actions beyond the intended scope  
- mitigation: Define authorization and tool-use boundaries

----- 5 Unverifiable claims
- The assistant may claim access, authorization, storage, or capabilities it does not have
- mitigation: Require capability and evidence checks before making such claims
----- Unclear uncertainty handling  
- Assumptions may be presented as facts   
- mitigation: Require explicit uncertainty and evidence labeling