In a Linux community discussion, users reported that the Gemini Flash model performs poorly in following instructions, failing to accurately execute verbatim copying tasks. For example, it incorrectly output “的核心技术壁垒” instead of “核心的技术壁垒”. Although users had explicitly provided feedback in the prompt to avoid such errors, the model persistently repeated the problematic behavior. While there was a small probability of outputting different content, it did not enter a complete infinite loop. In contrast, the Claude Haiku model performed more reliably, never exhibiting similar issues. This phenomenon reveals the reliability differences among AI models in prompt engineering, suggesting that users need to consider instruction accuracy requirements when selecting models. The discussion involved three posts from two participants, providing practical cases that serve as a valuable reference for readers concerned with AI performance and prompt optimization.
