Your question is Debug Negative Prompt Constraints. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
You're working on an LLM feature where the prompt includes a negative constraint, such as "Do not mention the word 'blue'." Most outputs follow the instruction, but some intermittently violate it. You need to figure out whether the issue comes from prompt wording, context interactions, decoding behavior, or something else in the prompt stack.
Describe your process for debugging a prompt that intermittently fails to follow a negative constraint, for example, "Do not mention the word 'blue'."