# Context Distraction Context distraction occurs when irrelevant or low-priority information in an AI agent's context diverts its attention from what actually matters. The model's attention mechanism treats all context tokens as potentially relevant, so noise competes with signal for the model's limited processing capacity. This is distinct from [[Context Bloat]] (too much context overall) and [[Context Poisoning]] (wrong information). Context distraction is about correct but irrelevant information pulling the model off-course. ## How it happens - **Over-inclusive retrieval**: RAG pipelines that pull in tangentially related documents, burying the actually relevant ones - **Unfocused system prompts**: master prompts that include instructions for every possible role and task, even when only one is active - **Context accumulation**: long conversations where early context (no longer relevant) still occupies attention - **Verbose tool outputs**: tools that return full documents when only a few lines were needed - **Kitchen-sink skills**: AI skills that load every possible piece of context "just in case" ## Impact Context distraction directly degrades output quality. The model's attention mechanism has limited capacity. When irrelevant tokens consume attention budget, the model: - Misses or de-emphasizes the actual task requirements - Produces generic responses because it can't focus on specifics - Follows outdated or inapplicable instructions that happen to be in context - Generates longer, less precise outputs as it tries to address everything it sees This connects to the principle that [[AI context is finite with diminishing returns]]. Past a certain point, more context actively hurts. ## Mitigation - **[[Progressive Disclosure]]**: load context incrementally, only what's needed for the current task - **[[Prompt Lazy Loading AI Design Pattern (PLL)|Lazy loading]]**: don't front-load everything; let the agent pull context on demand - **[[Receptionist AI Design Pattern|Receptionist pattern]]**: route to specialized agents with focused context rather than one agent with everything - **Context windowing**: actively expire or compact old context as conversations progress - **Focused retrieval**: tune RAG pipelines for precision over recall ## References - ## Related - [[Context Bloat]] - [[Context Engineering]] - [[AI context is finite with diminishing returns]] - [[Context Poisoning]] - [[Context Confusion]] - [[AI Context Rot]] - [[Context Hygiene]] - [[Progressive Disclosure]] - [[Prompt Lazy Loading AI Design Pattern (PLL)]] - [[Receptionist AI Design Pattern]] - [[Token Budget]] - [[Context Window]] - [[Personal Context Management (PCM)]] - [[Obsidian Starter Kit - Tutorial - Managing AI sessions]] - Operational guidance to prevent context distraction in OSK sessions