Работа
Division of labor: cheap filter → expensive think → agent execute → you decide
Burning the strongest model on simple yes/no judgments wastes money and attention.
Что болело
NextOrderAI (@NextOrderAI) argues the wasteful pattern is making the strongest model do the simplest judgments all day. Efficient play is division of labor: Sol for hard/deep thinking, Jev for mass screening/judgment, Grok Bot for concrete execution. Core logic: cheap models filter first, expensive models think next, the agent executes, humans only make the final call. Example: 10k leads — don’t run the strongest model on every row; Jev screens most out, Sol gets the worthy few, then the agent writes mail / updates CRM / follows up. What matters: what is worth thinking, what isn’t, and who does the work after thinking. Tip only. Not affiliated with xAI.
Что установить или создать
Screen cheap → think expensive → execute with Grok Bot → human final yes. Fewer tokens on “skip.”
Боты
Execute-lane bot
Runs only the concrete actions after cheap filter + expensive think; you keep the final call.
Плагины
В первый день не нужны. Следуйте шагам первого запуска.
Файлы
Optional routing card: which job class goes to filter / think / execute.
Рутины
На первом запуске их нет. Сначала пройдите один раз вручную.
Промпт
Это совет, а не системный промпт, который можно вставить. Автор его не опубликовал. Следуйте шагам первого запуска.
Первый запуск
- 1Read https://x.com/NextOrderAI/status/2103640327768973398 for the Sol / Jev / Grok Bot split.
- 2List one backlog (e.g. leads or tickets) and mark what is pure screen vs needs deep think vs needs execution.
- 3Run a cheap/fast model (or light pass) to drop obvious nos before any frontier call.
- 4Send only the survivors to a strong model for judgment, then hand concrete actions to Grok Bot.
- 5You make the final call on send/CRM/spend — agent count is not the same as system strength.
Барьер одобрения
Outbound mail, CRM burns, and spend still need your yes after the expensive think step.
Что поправить
Steal the three questions: worth thinking? skip? who executes after? Agent count ≠ system strength.
Источник
AI第二幕 @NextOrderAI · Исходный пост
Не связаны с xAI. Этот промпт написали не мы. Условия сторонних шаблонов по-прежнему действуют.
Пакет Pro
Нужна полная настройка, а не только идея?
В пакетах Pro — текст задачи, промпты, расписания и подводные камни для популярных сценариев.