Three things share the window: the system prompt, everything the user has sent, and everything the model has written back. They draw on one budget, so a longer system prompt leaves less room for the conversation, and every token the model writes is a token the next turn cannot use. Marking the three separately shows where the space went; what is left over is the room the next turn has to work in.
Paid members: the worksheet and its printable PDF are below ↓