What is context-window waste in AI coding?
The short answer
Context-window waste occurs when an AI model receives information that does not materially help it complete the current task. In coding workflows, this can include irrelevant files, duplicated source code, unnecessary dependencies, old context and repository content loaded primarily to discover where relevant code exists.
Waste, drawn
Each square is a file in a repository. The lit ones are the files that carry the answer to a single question. Everything else that gets read on the way to them is the waste this page is about.
Classifying context
Not everything sent into a model's context window plays the same role. Splitting it into rough categories makes it easier to see where waste is likely to accumulate.
Essential context
The code and information the model must have to complete the task correctly: for example, the function being modified and the type or interface contracts it must respect.
Supporting context
Related code that helps the model produce a better or more consistent result but isn't strictly required for correctness: nearby helper functions, similar patterns used elsewhere in the repository, or related tests.
Discovery context
Content loaded primarily to figure out where relevant code lives, rather than to complete the task itself: search results, directory listings, or files opened only to check whether they turn out to be relevant.
Irrelevant context
Content that doesn't materially help the task at all: unrelated modules pulled in by an overly broad scan, stale versions of files, or duplicate copies of code already present elsewhere in context.
The Context Efficiency Ratio
One way to reason about how much of a model's context window is doing useful work is to compare relevant context against total context supplied:
Relevant context tokens ─────────────────────── × 100 Total context tokens