The Context Window Is a Hard Ceiling | AI Learning Guide

The example is the part to hold onto. You paste a long report, add formatting rules, and ask a question. If the total runs past the window, the earliest content is what gets dropped, and that is often the section your question depends on. The formatting rules survive only because they came later. So the model is not ignoring you; those tokens are simply no longer available to condition on. The habit that follows is to treat the window as a budget: keep what must be present, cut the rest, and break long tasks into stages so only the distilled result carries forward.

0:00 / 0:00

Auto-play

The context window counts tokens across your instructions, pasted text, prior conversation, and the reply being generated. It is shared capacity, so every part competes with every other part.

What overflow looks like

You paste a forty-page report, add a detailed set of formatting rules, and then ask a question. If the total exceeds the window, the system may silently drop the opening pages. Your formatting rules survive because they came later, but the report section your question depends on may be gone. The answer can look confident and still miss the point, because the relevant tokens were never in play.

Treat the window as a budget. Decide what must be present, cut what is merely nice to have, and stage long tasks so only the distilled result carries forward.