I stopped babysitting my support bot when I added a reviewer agent prompt after every draft

Описание и применение

Добавление отдельного шага проверки (reviewer agent) в цепочку обработки тикетов улучшило надёжность поддержки, устранив ошибки в тональности, пропущенные данные и нарушения политики, которые не удавалось контролировать одним запросом к модели.

Задача

A single GPT call that tried to classify, draft, enforce policy, and check tone produced unreliable support replies that were technically correct but rude, overconfident, or missing key information.

Результат

Adding a dedicated reviewer step dramatically improved reliability, catching policy violations, tone issues, and missing information that a single prompt could not consistently handle.

Рабочий процесс

• 1. Classify the ticket type using a classifier agent.
• 2. Draft the reply with a draft agent based on the classification.
• 3. Review the draft with a reviewer agent that checks policy, tone, and completeness.
• 4. Escalate to a human if confidence is low or errors are found.

Доказательства

That extra review step caught policy misses, weird tone, and missing fields before anything went out.

Источник

One prompt was doing four jobs badly

Источник