Usage & Prompting
Context Window
The model's working memory: everything it can see at once in a single pass. What doesn't fit does not exist for the answer. Large windows allow whole contracts or codebases in one go, but attention is spread thin: what sits in the middle of long contexts tends to get lost. Selection remains more important than volume.
An example
A model with a 200,000-token context holds about 300 pages. So a team dumps the entire contract folder in and asks about one clause. The result is worse than expected: what sits in the middle of a very long context is measurably less well taken into account than what is at the beginning or end. The same question with the four pages that are actually relevant is answered reliably by the same model, and at one-fiftieth of the price.
A closer look
Everything the model considers for an answer has to fit: system prompt, history, attached documents, and the answer itself. When the window is full, the beginning falls out. The model does not point this out.
Why it matters
A large context window is an option, not a strategy. It costs linearly more money, disproportionately more compute time, and does not bring quality along. The reliable way to build remains: selectively retrieve what belongs to the question and pass along only that. The large window is then the buffer for cases where the selection needs more material than expected, not permission to skip selection. Anyone who reads provider claims about context length as a quality feature is comparing the wrong number.
Commonly misunderstood
A large window does not replace selection. In the middle of long contexts the hit rate measurably drops, and every additional token costs money and time.