Context confusion is the state in which excessive information, overly abstract instructions, or instructions that contradict one another leave a language model unable to judge which information to prioritize. Spec駆動開発におけるコンテキストエンジニアリングとCustom Slash Commandsのベストプラクティス names it as one of three problems that Context engineering is meant to address, alongside Context rot and Context poisoning. Its symptoms, as that post describes them, are results that vary widely from one run to the next and a greater tendency to hallucinate.
Usage
The post explains the cause in terms of the Signal-to-Noise Ratio (Code Review) of the context: when the information the task actually needs (signal) makes up a smaller share relative to unrelated information and redundant logs (noise), the model’s attention weighting stops working properly, so important information is under-weighted and unimportant information over-weighted, and output quality becomes unstable.
Its worked example comes from designing Custom slash commands: a single command that chained web research, document creation, guideline checking and correction was pulled off course by the large amount of noise from web search, which the author identifies as causing context confusion and context rot. The remedy the post proposes is to split such work into commands with a single responsibility each and to inject only the minimum information a task needs.