One bot manages six bots, sends only one message a day

Lauren Tan from the Grok Bot team has her own assistant setup, which Morlex organized into 12 files; the complete prompts are printed on the card.

The core structure is simple: she talks to only one bot, and that bot manages the other six. She receives just one message a day.

Most people approach configuring an AI assistant by finding an all-purpose bot and throwing everything at it. Lauren does the opposite. One bot only does routing and summarizing, while specific tasks go to six specialists who each handle their own area.

Among the 12 files, six markdown files define the behavioral boundaries:

prompt.md is her only point of contact. All requests come here first.

triage.md performs triage. Small tasks are completed on the spot; large tasks are forwarded to the responsible bot.

daily-summary.md sends one message a day, divided into only three categories: needs me, completed, blocked.

ask-first.md is the permission red line. Without her approval, no payments, no posting, no bookings, and no contacting anyone.

two-fails.md prevents infinite loops. If a bot fails twice, it stops and explains why.

quiet.md defines silence. If there is no news, no message is sent.

Under the bots/ directory are the specialists' division of labor: matcha assigns engineering work, and engineers run cloud agents inside it; dr. eggbot builds new bots and rereads chat logs every week; potato reads X mentions; suki handles travel; crumb tracks disk space.

This setup has been written into files; in essence, it turns management actions into configuration. Triage, reporting, authorization, circuit-breaking, and silence are all things a chief of staff does in a human organization.

Someone asked Morlex: if you talk to one bot and it then instructs other bots, the same task consumes resources twice—what is the logic behind that?

This question is worth answering. If only one bot does everything, its context becomes increasingly polluted. Engineering task logs, travel booking confirmation numbers, screenshots of X mentions—all pile up in one memory. With every added task, previous tasks create another layer of interference.

After domain separation, each bot has its own independent context and memory. The chief does not do the work; it only decides what should go to whom and which matters need the boss to make the call. A chief of staff does not write code for engineers; it manages the boss's attention.

A month ago, Grok Bot switched its backend model to Claude Opus 5.5. Musk posted on October 7 that SpaceX would route tasks to the best model, and the first name was Opus 5.5. Lauren later confirmed that all bots are powered by Opus 5.5 and can also spawn Cursor cloud agents.

This means that when matcha assigns engineering work, it can launch Cursor agents to run specific tasks, while the Opus-powered bot is responsible for planning and review. In his long post, Morlex calls this 'the coding bot becomes a manager.'

But one question remains unanswered. A user asked: if the bot doesn't tell you which model it used, how do you know it is really using Opus 5.5?

Currently, there is no public model identification mechanism. Another user said token consumption is too fast. Morlex's article also mentions that Opus 5.5 is considerably more expensive than Grok on the API, and although Grok Bot charges by subscription rather than by token, whether Opus-driven tasks will burn through quotas faster has not been stated officially.

The value of Lauren's setup is not in the six bots themselves. It is a template: writing into explicit rules what should be done automatically, what should stop and ask, how to stop after a failure, and don't bother me if there's nothing going on.

Most AI assistants do not lack capability; they lack boundaries. Between a bot that dares to do anything and a bot that knows when to stop, the latter is more worth putting into a daily workflow.

The complete prompts are on Morlex's card. Paste them into Grok Bot, and you can have your own chief tonight.