Training an agent is not prompt writing. It is five stages, each with an artifact that has to exist before the next one starts: discovery, knowledge base, tuning, evaluation, production. Skipping evaluation is why most pilots sound impressive in a demo and fall apart in week three.
Discovery is two hours of your real recorded calls plus the questions your front desk answers most. The knowledge base is the artifact: service menu, pricing tiers, hours, service-area boundaries, escalation rules, and the FAQ set, chunked and embedded so retrieval returns the right paragraph rather than a plausible one.
The system prompt is written from your transcripts, not from a template. Cadence, the way you handle price objections, what you say when you cannot help — those come out of the recordings. Edge cases get explicit handlers: hostile caller, Spanish speaker, out-of-scope request, someone asking for the owner by name.
A fixed set of scenarios the agent has to pass before it sees live traffic: the emergency that must escalate, the booking that must not double-book, the question it must refuse to answer, the caller who changes their mind mid-sentence. Failures are logged with the transcript, fixed, and re-run. This stage is why the go-live date is 21 days rather than 3.
Live traffic produces escalations, and escalations are the next week's training inputs. Weekly you see call volume, booking rate, escalation rate, average handle time, and the five questions that most often needed a human. Those five become prompt or knowledge-base changes. The agent gets measurably better in month two, which is the whole point of the loop.