How Claude Code Set Up 16 AI Assistants for My Business in 18 Minutes,and the 4 Rules From the 5 Days of Planning Before It

Claude Code, my AI coding assistant, set up 16 AI assistants for my business in 18 minutes. That was the easy part: the 5 days of planning before it produced 4 rules that kept the assistants from going wrong, and the habits behind them work for any new helper, human or AI.

What a Grok Bot is

The assistants are Grok Bots, helpers inside xAI's Grok app. They answer me from my phone. They run in the cloud, and they can also run tasks on a computer I choose.

That mix is why I picked them. Before this, I had tried two other agent setups, and neither worked well. The hard part was getting the agents to run in the cloud and answer my messages in real time, so I wasn't tied to my desk. Grok Bots run in the cloud and reach me on my phone. What I still needed was a plan for what they were allowed to touch.

The first plan asked me to approve everything

On October 1, I asked Claude Code to plan the team with me. Its first plan needed my approval on nearly every step.

I pushed back. If I have to approve everything, even the trivial things, the bots defeat their own purpose. The whole point was fewer things to approve and more free time.

I also asked for a different kind of safety. A bot that reads my email can, sooner or later, read an email written to trick it. I didn't want a bot that is too fenced in to do its job. I wanted a bot that is better at spotting a bad request hiding in an email, and that still does the job.

Over the next few days we went back and forth, and three other AI coding tools reviewed the plan. We landed on 4 rules.

The 4 rules

1. One computer does the work

The bots may run commands on 1 separate Mac. I asked for a new machine with no saved keys lying around on it for a bot to find. My other Macs are set to "Never allow," so a bot can't run anything on them.

2. One rulebook

Every bot's instructions sit in 1 shared folder on that Mac, and the bots reread it before every task.

The copy I pasted into the Grok app is short. It mostly tells the bot where the real rules live. The folder is the real copy, and that Mac pulls the latest version every 5 minutes. When I change a rule, every bot has it within 5 minutes, and I never paste anything into 16 chats again.

3. Rehearse first

Before any bot saw my inbox, Claude Code played each bot on my laptop, in a sandbox that caught every message before it could go out. Texts, emails and saved records all landed in a holding pen instead of reaching anyone. Afterward, a separate check looked at my real sent mail and messages to confirm nothing had slipped through.

The first rehearsal found 104 problems. 18 were quirks of the test setup itself. 4 were environment issues that had already been fixed. 6 were decisions only I could make. And 76 were mistakes in my own instructions.

That last number is the one I keep coming back to. Most of what would have gone wrong was not the AI misbehaving. We fixed all 76 before any bot went live.

New bots also start out on the safe side. They draft messages and wait for my yes in that bot's own chat before anything goes out.

4. Make it read back

On October 6, Claude Code clicked through the Grok app the way I would and created the bots, one after another.

Each bot was asked to reply with the character count of the instructions it had saved. Claude Code compared that number with the length of the file it had pasted. All 16 matched.

That one check tells you whether the helper got the whole message, not most of it. A rulebook that was cut off halfway looks fine until the day the bot needs the missing half.

The team has grown since

By this week I was running 23 agents.

I manage them day to day with Antigravity, a second AI coding tool. On October 30 I gave it my daily to-do list, and it turned the items into scheduled jobs for the bots. The team went from 18 to 20 bots that night.

The next night it checked all 20 bots and found 9 whose schedules were missing or paused. The same habit as rule 4 paid off again: ask the system to show you what it actually has, instead of trusting what you meant to give it.

The 4 rules kept the bots safe. They didn't keep them cheap.

More bots on more schedules is also how I used up a week of AI budget in 2 days, which I wrote about in my last post.

Safety and cost are separate problems. Nothing in these 4 rules asks how often a bot should wake up, or whether it has any work to do when it does. If you are planning your own team, plan for both.

What you can copy, without writing code

You don't need to write code to do this in your business. The same 3 habits work for an AI tool you rent, a virtual assistant, or a new hire:

1. Put your helper's rules in 1 document. One place to read, one place to change.
2. Test it on fake messages before a real customer sees anything. Expect most of the problems to be in your instructions.
3. Have it repeat back what it was told before you trust it.

What's the first rule you'd write down for a new helper starting Monday?