What the Model Sees (one string, every turn, from the top)

The first board of the Building with LLMs run, standing on the serving run before it: The Two Clocks (TTFT, TPOT, and why goodput is the number) priced a request from the machine’s side, Prefix Caching (the same start, computed once) showed why its first tokens can be computed once and reused. This board turns to the app’s side: four chat bubbles on the left, the one long string that actually went over the wire on the right.

Before you start. A model reads and writes in tokens, chunks of text about four characters long, and it is paid per token it reads and per token it writes. Reading a prompt is fast, because every token is read at once; writing is slow, because each token waits for the one before it. A tool call is the model writing out a request like lookup_order(4471), which the app’s own code runs before pasting the answer back into the text. This lesson looks at all of that from one level up, as the single piece of text rebuilt from nothing and resent in full on every turn.