see that progress had been made, and declare the job done
这是大语言模型常见的“过度乐观”陷阱。模型倾向于迎合用户的完成预期，而非客观审视实际进度。通过强制读取结构化的feature list，是用外部状态锚定来对抗模型的内在偏见。
1 Matching Annotations
- Last 7 days
-
www.anthropic.com www.anthropic.com
-