In reasoning mode, the chain of thought (CoT) tends to leak the user's instructions. What are the best practices to mitigate this? I noticed that when using the exact same prompt with DeepSeek-R1, the instructions remain hidden.
· Sign up or log in to comment
