Figure 7.1 The anatomy of an assembled context. What the model reads on a mid-run
pass is a single stacked window, but its slices have four different authors:
you write the standing instructions and the required output shape, the user
writes one request, the model writes the running conversation, and the harness
fills the rest—memory, retrieved documents, tool definitions, and the
often-dominant tool results. Prompt engineering owns the accented slice; context
engineering answers for the whole desk.
