Prompt
front-desk
Reads one inbound shop message, classifies the intent, extracts the booking fields, drafts a reply under hard rules, and decides whether a human needs to take over. Opt-outs never reach the model — a deterministic pre-check handles them first. Every prompt version is scored against the same hand-labeled gold set below.
UNDER TEST ▸ The front-desk classify+extract+draft prompt, tested across versions against the same 30 hand-labeled inbound messages (1 handled by the deterministic opt-out pre-check, 29 by the model).
scored on the gold set (the 29 cases v3 was tuned against)
FIG.3 · front-desk eval / rev.A. v3 was iterated against the gold set, so its 100% is a training-set number. The held-out line above is the real test. New messages, labels frozen before a single one-shot blind run, never used for prompt iteration. Its result stands as published, whatever it says.