Is your AI team stuck in hand-offs?

Why my AI team kept stopping to ask

From October 5 to 8, my AI setup ran 56 threads on my PC. Between me and those threads, work changed hands 293 times. And 32 times, a tool call stopped on an "Allow?" prompt and waited for me.

Here's why. I'd message a coordinator in the cloud, and it relayed every task to a fresh thread on my PC. Claude Code won't give full permissions to threads started from a phone or the Projects app, so those threads kept asking "Allow?".

Nothing was broken. There were just too many hops between me and the work.

The fix: one boss chat

Now I talk to one chat. I named it Steve. It's a single, long-running Claude Code chat on my PC, and it's the only one I message.

There's no cloud coordinator relaying each task anymore. I ask Steve, and Steve decides who should do the work.

How each ask gets routed

Steve sorts every ask into one of four lanes:

• Quick question: Steve answers it right there.
• Quick fix: one worker does it.
• New build: a Gatekeeper and a worker take it.
• Deploys, money or messages: Steve waits for my OK first.

The last lane matters most. Anything that goes live, costs money or sends a message stops and asks me before it happens.

The Gatekeeper checks the work and sends it back

A new build doesn't go straight from the worker to me. A Gatekeeper sits in between: a stronger model (for example Claude Opus) whose only job is to check.

It doesn't take the worker's word for it. It checks the work itself: it runs the tests, reads the changes and looks at the live result. If something is off, it sends the worker back with specific fixes. It does that for 3 rounds at most, then hands Steve a verdict.

Two real examples:

• The first build to go through the Gatekeeper passed in 1 round, with 120 tests passing.
• This video took 2 rounds. The Gatekeeper sent round 1 back for polish.

A side pane shows every job

With a few jobs running at once, I need to see them without asking. So I keep a side pane next to the chat. It lists every job under Waiting on you, Working and Resolved.

When something needs me, it appears under Waiting on you with one-click answer buttons. I tap an answer instead of typing one.

What you can copy tomorrow

You don't need my exact setup. The idea has three parts:

1. One chat to talk to. Tell it to sort each ask: answer it, hand it to one helper, or send it through a checked build.
2. A checker for anything you build. A second, stronger AI whose only job is to check the work and send it back.
3. Your OK before anything that can't be undone. Going live, spending money and messaging people all wait for you.

Below is a general Gatekeeper prompt, rewritten so it works for any kind of job. Copy it and paste it as the first message of a new chat, or into your app's custom instructions or project instructions.

You are my Gatekeeper. I give you one job to get done right. You hand the hands-on work to a worker, check the result yourself, and give me a verdict I can trust. You never skip the check, and you never take a step I haven't approved.

1. Read the brief
Start from my brief: what I asked for, what I'm really after, what would disappoint me, and what "done" looks like. If any of those is missing, ask me before the worker starts.

2. Hand the work to one worker
Give the worker a complete brief: the goal, where to work, only the background this job needs, what it may change, and where it must stop and ask. Use a lighter, cheaper model for routine work. Use the strongest one available for hard builds (for example Claude Opus). Only use models this app actually offers.

3. Check everything yourself
"Done" isn't proof. Check it with your own eyes, in a way that fits the job:
- Code: run the tests and read the changes.
- Settings: look at the saved value after a reload.
- Websites and documents: open the finished thing and look at it the way a customer would.
- Numbers and dates: compare them against the source.

4. Judge it against what I wanted
Passing the checks isn't enough. Ask whether this is what I meant and whether it would disappoint me. If it's cluttered, confusing or sloppy, send it back even though it technically works.

5. Send it back with specific fixes
If anything falls short, start a fresh worker for the next round. Give it the brief, the last report, your exact corrections (what is wrong, where, and what "fixed" looks like) and where the work lives. Never just say "try again".

6. Stop after 3 rounds
If it still isn't right after 3 rounds, stop. Tell me exactly what is still wrong and let me decide.

7. Never take a step I haven't approved
Reversible work is fine. Always ask me first before anything that can't be undone or reaches outside this job: publishing or deploying, spending money, deleting things, or messaging other people.

8. Give me a verdict
End with short, plain lines: passed, failed or blocked; how many rounds it took; what you checked and what you found; whether it matches what I wanted; and anything still waiting on my OK.

How to use it in your AI app

• Apps with helpers built in (Claude Code, Codex and other agent tools): paste it as the Gatekeeper's instructions along with your brief, and let it start the worker itself.
• Chat apps (Claude, ChatGPT, Gemini): use one chat as the Gatekeeper and a separate fresh chat for each worker. Carry the briefs and reports between them by hand.
• Model names: swap in your app's own lighter and stronger models if it doesn't offer Claude Opus.

Start with one real job you'd normally trust on the first try, and see what the checker catches.

Where does your AI setup get stuck?

Is it the hand-offs, the permission prompts, or trusting the result? Tell me in the comments on @autom8everything. For more like this, see all my AI tips. And if you'd rather have a setup like this built around your own business, book a free 30-minute call and we'll talk it through.