Transcript Format Selection

Choose how a completion model's conversation transcript is written on the page — freeform, script-style, markerless, or structured tags.

Writing an advice conversation document for a completion model requires choosing a transcript format — nothing enforces one the way a chat API's template would. Four formats, each addressing a weakness of the previous:

Freeform text — flexible for inserting content between quotes, but hard to assemble dynamically and reliably once there are many elements to interleave.
Transcript/script format (Me: … / Assistant: …) — easy to assemble programmatically, but weaker for long or formatted elements, such as indented source code, which doesn't sit naturally after a speaker label.
Markerless format — works well with formatted text and long pasted content (e.g. emails), but the model can struggle to track who is speaking, and the application can struggle to detect where the model's response ends.
Structured format (e.g. XML-tagged turns, see structured document format) — clearly indicates who is speaking and when they finish, at the cost of more scaffolding tokens per turn.

Nothing prevents the prompt engineer from writing the assistant's lines too — an application of the same inception technique used to transition into an answer: writing from the assistant's perspective, as if it had already asked a clarifying question, frames the next real user turn as responding to that question, and ensures the actual completion opens with an answer rather than another round of clarification.

Sources

Prompt Engineering for LLMs (John Berryman), ch. 6