Fable 5 遇到瓶頸 —— 而問題出在你的盲區
anthropic 的 fable 5 效能現在受限於使用者的盲區而非模型智慧，這要求開發者在進行提示詞工程前必須先識別「未知的未知」，以避免產生平庸的輸出。
shihipar explains that fable 5 shifts the bottleneck from the model to the user, specifically their 'unknown unknowns'—factors never considered that could derail implementation. he suggests a 'blindspot pass' technique where users explicitly ask the model to find gaps in their plan before coding begins, as rigid instructions prevent course correction and vagueness leads to industry defaults that don't fit the specific task. this discovery process implies that achieving optimal results with frontier models now requires meta-cognitive prompting strategies to force the model out of rigid instruction-following modes.