Fill the window

The context window has a token limit. Both the supplied material and the growing reply need room. It can be large, but it cannot grow forever.

Press add a turn. The app instructions are added again at the top of every request. Below them, the conversation grows one user-and-assistant turn at a time.

This demonstration uses one common way of making room: once the example window is full, the app removes the oldest conversation turn before sending the next request.

Watch the early request to "keep answers short." Once that turn is removed, the model cannot use it. The words are no longer in the input.