The orchestration playbook.

One agent, one job, its own files.

• Buildwrites the codeClaude Code
• Testtaps through it like a userCodex
• Artmakes icons and screenshotsChatGPT images
• Reviewreads the codeCodex

I once put two agents on the same files. They kept undoing each other, like two coworkers fighting over the air-con remote, and it was slower than one agent. Now each agent gets its own files.

My review trick: tell Codex that Claude wrote the code. It gets much stricter.

How the jobs flow.

I set the North Star (my one goal number) and write the plan. Grok Bot is my chief of staff, and it hands each agent one job. The agents check their work, then my engineering Grok bot reviews and merges it. Anything that fails goes back for a fix.

The rule that keeps them on track.

My Bottleneck Killer template runs this every day.

1. 1Find the one bottleneck.The one thing holding my app back most.
2. 2Plan how to kill it.Automate it, or get rid of it.
3. 3Everything else waits.A stop list: what I don't build yet.

When the bottleneck is technical, it sends the work to my engineering bots. Get the template. You need Grok Bot to run it.

Guardrails.

Approving every step is too slow. Full access is too risky. As I posted in March: “Even if you have tight permissions on your AI, there's always a way around it.”

My rule: use auto mode, where an AI checks each command first. It's faster than approving everything, and safer than full access. I still check the work.

What I check instead of unit tests.

I don't use unit tests on any project. I check these instead:

• A test agenttaps through the app like a user
• Analyticsshows where people stop
• Crash logsshow what broke
• My reviewI read the risky changes

A loop that runs without me.

On Air Fry Pro, an agent runs the same loop every week. It reads the traffic, picks small changes, ships them, checks if traffic moved, and keeps notes for next week. Visitors went from 1,103 to 1,662 a month (DataFast, Aug 14 to Sep 10 vs Sep 11 to Oct 8).

Start here.

1. 1Pick one small app.
2. 2Answer the six questions in plan.md.
3. 3Tell one agent: read plan.md, then do one job.
4. 4Check it on your own phone.
5. 5Add a second agent only for a second job, on its own files.
6. 6Every day, kill one bottleneck. Everything else waits.