Emitted protocols: the model's structured output channel

The most powerful prompt-architecture idea is to define a structured format that the model emits and your code parses deterministically. Rather than trying to extract meaning from free-form prose ("did it cite anything?"), instruct the model to append a strict, machine-readable block, then parse it.

Examples of emitted protocols:

• A trailing JSON block of citations, each mapping an inline marker to a source, page, and verbatim quote.
• Inline tagged values like [[Yes]], [[USD]], or [[page:3||quote:...]] that your code reliably extracts.
• A marker convention that triggers behaviour (a message prefix that means "apply this template").

The pattern is always: the model produces structure, your code consumes it. This turns the model from a prose generator into a structured-data source you can build features on (clickable citations, typed spreadsheet cells, triggered workflows). Define the protocol precisely, give an example, and parse it with a simple, deterministic parser (a regex or JSON parse), not another model call.

When you stream output to a user, remember to hide the protocol from the visible stream: buffer just enough to detect the block boundary so the user sees clean prose while your code still gets the structured trailer. (Chapter 8.)