Article

The Standing Rules My Grok Bots Share, and the One I Retired

By Austin Hoopes · Published November 4, 2026

Most of what my Grok Bots know is specific to their one job. A handful of rules apply to all of them. I wrote each one down once, in the memory every bot shares, so I never have to teach it bot by bot.

None of these rules are clever. Each one exists because something small kept getting in my way. And one of them I’ve already retired, which turned out to be the most useful part of the exercise. About a four-minute read.

Why shared rules at all

Each bot I run has its own job description. That’s where the work lives: what it does, what it never does, when it stays quiet. (More on that in What is a one-job Grok Bot?)

But some preferences aren’t about any one job. They’re about me. How I read things, how I approve things, what I want to stay away from. Putting those in thirty job descriptions means thirty places to forget one.

So they go in shared memory instead. One fact, every bot. When a rule changes, I change it in one place. (For how bots keep and share what they know, see Agent memory for AI fleets.)

Rule one: write like a person

Every bot writes to me in normal sentence case, with real punctuation and grammar. Even when I type to it in lowercase with no punctuation, which is most of the time.

It sounds fussy. It isn’t. I type casually because I’m in a hurry. I read what comes back because I need to act on it. Those are different jobs, and the second one goes better when the text is clean. The rule overrides any bot persona that says otherwise.

Rule two: one tap to copy

A lot of what my bots hand me is text I’m going to paste somewhere else: a post, a reply, a caption, a message.

So whenever a bot gives me something to copy, it numbers the options, Option 1, Option 2, and puts each option’s exact text alone in its own code block. Notes and explanations go outside the blocks.

The point is the copy button. One tap, and I have exactly the text and nothing else. No trimming a stray “Here’s a draft:” off the front, no deleting a note I didn’t mean to paste.

Rule three: my yes travels

I talk mostly to one front-door bot, and it passes work to the specialists. (That setup is in What Running Thirty Grok Bots Actually Looks Like.)

The catch was approvals. I’d say yes to something in the front-door chat, the request would reach the specialist, and the specialist would ask me to confirm it again in its own chat.

Now the rule is that when one of my two coordinating bots relays a decision I made in its chat, every other bot treats it as mine and acts on it. That covers things like page copy, fixes, and changes to a bot’s setup or routines.

It has limits, and the limits are the important part. A relayed yes doesn’t cover sending an email or message to someone outside, spending money, deleting things, or any decision I haven’t actually made. Those still come to me directly. The rule carries a decision. It doesn’t create one.

Rule four: hands off my name on X

Bots can read X all day for research. They never post, reply, like, or send a message there for me unless I ask for that specific thing. Anything that goes out under my name waits for me.

This one used to be stricter, which brings me to the rule I retired.

The rule I retired

For a while, no bot was allowed to touch X at all without my approval, even to read. Before asking, it had to tell me what the call would cost.

That rule had a good reason. The X connector I used then drew from a paid balance of API credits, and I didn’t want any bot spending it quietly.

Then Grok Bot got its own built-in X tools for reading posts, profiles, and search. One of my bots ran a search through them and checked the balance before and after. It didn’t move.

So the reason for the rule was gone. In late October I retired it. Bots now read X freely for their own jobs, without asking and without a cost estimate.

The part about acting as me stayed, because its reason didn’t change. A post under my name is still a post under my name, free or not.

What retiring it taught me

Every rule I give a bot now comes with its reason attached. Not for the bot’s sake. For mine, a month later.

A rule without a reason just sits there. Nobody questions it, and it keeps costing a little friction long after it stopped protecting anything. A rule with its reason written next to it can be checked: is that still true?

When the X reads became free, I didn’t have to remember why the approval rule existed. It said so. That made it easy to keep the half that still mattered and drop the half that didn’t.

If you want to try this

• Notice the repeat. If you’ve corrected the same thing in two different bots, it’s probably about you, not the job. That’s a shared rule.
• Write it once. Put it where every bot reads it, not in each job description.
• Attach the reason. One sentence on why it exists.
• Draw the edge. Say what the rule does not cover. Mine are mostly limits on sends, spending, and deletes.
• Check the reasons now and then. When one stops being true, retire that rule or the part of it that no longer applies.

For the bigger safety picture behind rules like these, see Keeping AI agents safe.

What’s a rule in your own setup that’s still running after its reason left?

← All articles