Prompt Engineering Playwright Metaphor

The visible user-assistant conversation and the actual model transcript differ — the transcript is a script with several collaborating authors.

A key source of confusion when building chat applications: the human user's visible conversation with "the assistant" is not the same thing as the application-to-model ChatML transcript actually sent as the prompt. The transcript can contain information the human user never sees — for example, when a user asks "How should I test this code?", the application must infer what "this code" refers to and fabricate supporting user or assistant turns containing the relevant code snippet before sending the transcript onward.

A useful metaphor: a theatrical play, with characters, a script, and multiple collaborating playwrights.

Characters are the ChatML roles — user, assistant, system, tool (names vary by provider).
Script is the prompt: the transcript of character interactions working toward solving the user's problem.
Playwrights, plural, jointly write that script:
The prompt engineer determines the overall prompt structure and boilerplate text — system prompt architecture and prompt anatomy.
The human user introduces the focal problem or theme.
The model itself fills in the assistant's speaking parts, though the prompt engineer may also pre-write portions of assistant dialogue via inception.
External APIs supply additional injected content — a documentation search tool contributing retrieved passages, for instance, via function calling.

Extending the metaphor: the prompt engineer is the lead playwright and showrunner, responsible for the overall shape and tone of the interaction — not writing every line, but directing what kind of play gets performed and keeping the other playwrights' contributions coherent with each other.

Sources

Prompt Engineering for LLMs (John Berryman), ch. 3