Not allowed to say hi

Grok Bot has been out for a month. Most of what I have read about it assumes a running business: a sales bot, an ops bot, an invoice bot, a chief of staff on top. I'm not there yet.

The thing I'm building is called Bravoure. Engineering teams keep adding CI jobs, bots, and coding agents. I'm building the part that keeps track of what is still running, what it costs, and whether anyone reads what it produces. My weeks go into that and into talking to the people who live with it.

So I set up three Grok bots. A week in, I still change something every day.

Scout finds named people on platform and DevEx teams and hands me names and links. It does not write the message and it does not bring up what I'm building.

Clerk reads my inbox and turns anything that looks like a research reply into a note I can paste and a draft I can send. It never sends. Its first run found nothing and said nothing, which is the output I wanted. It then widened the window on its own to find something. That became a rule the same evening: 24 hours means 24 hours.

Freeze compares the live site and the GitHub App listing against a list of allowed claims that it reads and does not edit. Anything new comes back with its source. The first check ran on the wrong bot, because I sent it to the wrong bot, and flagged one thing: my own email address on my own contact form.

What I changed from the articles I copied:

Every bot reads one file before any task. Who I am, what I'm building in two lines, what no bot may ever do. Their own notes stay their own. A fact about me that lives inside one bot is a fact the other two will get wrong.

Only one bot is allowed a schedule, and it is the one whose check is a file it cannot edit. Freeze diffs against a list I own. Scout and Clerk are judged by me, so they run when I ask. Without that, a schedule is just a clock. It pays a model to agree with itself every morning.

I am the gate on two things: who I talk to, and what gets sent. Nothing else waits on me.

Corrections go into the bot's own description as a rule, "always X, never Y". The first evening produced three.

There is no chief of staff. No bot writes code; that still goes through Cursor, where the tests are, on a machine none of them can log into.

None of this is a conclusion about Grok Bot. It is one setup, a week in, written so someone at the same stage can try it tomorrow, and some of it will be wrong by next week.

I thank these people for the inspiration!

@kingwilliam_ on memory, @0xCarnagee on the roster, @0xRafy on loops, @hanakoxbt on graphs.