A thought-provoking blog post from a Chinese developer introduces the 'SKILL First Law': large language models reset their cognitive state with every session, forgetting previously hard-won corrections and treating all information equally. This ephemerality causes AI to produce low-quality, manual-like outputs that lack depth. The author argues that we need to 'rub' or trace the AI's transient cognition to create comparison baselines, enabling it to recognize its own blind spots. For developers and engineers building AI-dependent workflows, this insight is crucial. It suggests that effective prompt engineering must account for session memory loss, perhaps by injecting external context or using iterative feedback loops. The concept has implications for AI-assisted coding, documentation, and research, where consistency and depth matter. While not a technical paper, the idea is novel and practically useful, offering a fresh lens on a common frustration with LLMs.
A new mental model reveals how LLMs forget their own corrections across sessions, degrading output quality. The 'SKILL First Law' offers a practical framework for developers to improve AI-assisted work.