Token density

A rambling seller phone call and a tight bullet list can carry identical information at wildly different token cost, and bloated context doesn't just cost money, it makes answers worse. Bury the one detail that matters (the roof is 2 years old) in noise and the model's recall drops. The same thing happens to a whole session. Early on, the model is in its smart zone: sharp, following every instruction. As the session fills, quality slides into a dumb zone, slowly and with no warning: it forgets an instruction from earlier, repeats a fixed mistake. That slide starts well before the window is full. Plan around the smart zone: one task per session, and dense context so each task gets the sharp part.