GenAI Field Guide

Under the hood (just enough)

The context window

The context window is the maximum number of tokens the model can attend to in one call. The whole assembled prompt plus the reply must fit inside it.

≈2 min
1,000,000-token window Your document 400k tokens Conversation + reply About 570k tokens left 30k preamble + tools also take space
Usable space = window − preamble − tools − memory − files. A 1M window with a 30k preamble and a 400k document loaded has about 570k left, and the reply must fit in that too.

Three facts that change how you work

Fact What it means for you
Every turn re-sends everything. The call is stateless; turn 40 re-sends turns 1–39. Long chats get slow and expensive (vendors cache the unchanging prefix to soften this). Start a fresh chat for a new task.
The reply goes back into the context, appended verbatim, then your next message after it. The transcript is the context. Nothing is "stored in the model." If it isn't in the transcript or a file, it isn't anywhere.
The reply must fit too. If you loaded most of the window, ask for a shorter answer or split the job.
Turn 1 Your message Turn 2 Turn 1 + reply + new message Turn 3 Turns 1–2 + replies + new message
The whole transcript travels again on every turn. Turn 40 re-sends turns 1–39.