context-window-management

Large language model performance degrades during long conversations because the context window becomes overloaded with excessive instructions and tool results, causing a drop in output quality and focus. Managing context window utilization through instruction budgeting and state offloading prevents this degradation.

To manage AI context windows for large refactoring tasks, apply context budgeting to plan instructions, offload large tool results and file contents to temporary files, and save session state to markdown artifacts to keep the AI focused and effective.

Context budgeting optimizes AI performance by planning and allocating instructions across different sources to stay within model limits. This ensures the context window is not overloaded, maintaining consistent output quality during complex multi-step tasks.

You should restart an AI session when performance degrades due to excessive context and instruction count thresholds are exceeded. Before restarting, save the current conversation state to a markdown artifact to preserve progress and maintain continuity.

The best way to offload large tool results when working with large language models is to save intermediate results and session state to files. This reduces reliance on volatile context and frees up the context window for essential instructions.