You can find the one job AI should be running for you by Friday, using the scale Anthropic built to grade its own work.

The number that got the headlines: Claude now leads 26% of Anthropic's own AI research. In February it was under 1%. Over 90% of that work has Claude at least collaborating. Thirty thousand Claude agents run at once on their internal platform. Nothing in there runs fully on its own.

The number I care about is smaller. Monitors review every single action those agents take before it happens, and block about one in 47,000. That's the whole setup in one line: let it lead, watch everything, veto almost nothing.

The level where your time comes back

Anthropic grades every task on six levels, AL0 to AL5. The two that matter for you:

Collaborates (AL3). The AI does a large chunk of the task, you steer it closely. Most people who use AI daily live here on everything.

Leads (AL4). You give it the goal, it takes the task to done. Then you check the result.

The gap between those two is where the 90% is. It isn't another tool or a better model, it's moving one task from "I steer every step" to "I check the result".

Now a quick word from today's sponsor.

Stop Dabbling With AI and Start Earning With It

Ready to stop using AI as a search engine and start using it as an income engine?

The Hustle's "200+ AI-Powered Income Ideas" is your free playbook for turning the most overhyped technology of our time into actual cash.

Inside you'll find:

  • 200+ real, vetted ways to generate income with AI, spanning freelance services, digital products, content creation, and emerging markets

  • Actionable strategies built for non-engineers, so your technical background (or lack of one) won't hold you back

  • Ideas aligned with where the market is actually heading, not where it was two years ago

Subscribe free today and unlock the full guide. The people already cashing in aren't smarter than you, they just started earlier.

Right. Back to it.

Your week on the scale

Take ten minutes on a Sunday. List the jobs you did with AI last week. Next to each one, honestly: did it collaborate, or did it lead?

My guess is almost every line says collaborate, and one or two are ready to move up. The tell is a job where you've corrected the same thing three weeks running and it hasn't been wrong since. That's a job you're still steering out of habit.

Pick one. Not three. One. Then move it up a level with this, in whichever AI you use:

This is a job I do with you every week: [describe it in two lines].

So far I've steered every step. From now on I want to hand you the goal and check the result. Tell me:
1. What you need from me once, up front, so you can run it without asking each time.
2. Where you'd still stop and ask, and why.
3. What a finished result looks like, so I can check it in under two minutes.
Then run it.

The second answer is the one to read twice. If it says it'd never stop and ask, it hasn't understood the job.

Watch everything, veto almost nothing

Anthropic's monitors see 100% and block 0.002%. Copy the shape, not the scale. When a job moves to "leads", you still read every result for the first month. You'll change almost nothing. That's the point, and it's how you know it was ready.

What you don't do is let it lead and stop looking. Anthropic runs 30,000 agents and still flags 50 transcripts a week to a human. If they still look, you still look.

Make it stick

Write the job's brief into the place your AI reads first: a Claude Project, ChatGPT Custom Instructions, a Gem. Goal, what done looks like, where to stop and ask. Then next Sunday, one more job.

That's one job moved from steering to checking. An owner has twenty of them, and the ones that move first are the ones nobody wanted anyway.

If you own a business and want the boring work off your plate, that's what Run On Claude is for. Every repeat job you'd hire for, set up on Claude so it runs on autopilot, with the install files. You keep the decisions. Owners only. $147 a month until member 50.

Quick one.

Which job did you move up a level? Hit reply with the one line. I'll share the best ones next Sunday.

Talk soon,
Zephyr