The guy who built Claude Code doesn't prompt anymore
Boris Cherny runs Claude Code at Anthropic. He literally built the thing. And on stage recently he said something that broke the internet:
Read that again. The person who built one of the biggest AI tools on earth doesn't sit there typing prompts. He builds loops that prompt the AI for him. He even uninstalled his code editor. He writes loops and lets them run.
He said this is the shift we're all going to make this year. So if you're still typing into ChatGPT, fixing it, typing again... this one's for you.
Prompting vs loops (what's actually different)
Here's the old way. The way 99% of people use AI right now.
You write a prompt. "Build me a website." The AI gives you something. You look at it and go, eh, the text is messy. So you type "fix the hierarchy." It kind of does it. You type "make it pop more." Again. And again. An hour later you basically built the thing WITH it.
That's prompting. You are the glue between every step. One shot at a time. Nothing moves unless you push the next button.
A loop flips it. You set it up once. Then the AI does the work, checks its own work, fixes what's weak, and goes again on its own until it's actually good. You walk away. You come back to the finished thing.
Prompting needs you every step. A loop runs itself. That's the whole difference.
What a loop actually is (3 simple parts)
A loop sounds technical. It's not. You're just telling the AI three things up front instead of one.
Then it runs the cycle by itself: Write a draft. Score it. Find what's weak. Fix it. Score again. Keep going until it hits your bar. Then hand you the best version.
The magic isn't a better prompt. It's that the AI grades its own homework and keeps going until it passes.
The part everyone skips (it's the most important)
Everyone clipped the "stop prompting, write loops" line. Almost nobody caught the part that actually matters.
Boris said the single most valuable thing you can do is give the AI a way to check its own work. A score. A test. A checklist. Something that answers "is this actually good yet?" He put the quality jump at 2 to 3 times.
Here's why. A loop with no check is just a faster way to make junk you still have to fix yourself. The loop is the easy part. Anyone can tell an AI to "keep going." The real skill is writing the check that decides when it's done.
Writing the loop is the headline. Writing the check is the job.
So when you build a loop, the first question isn't "what do I want." It's "how will it know it nailed it." Answer that and you're ahead of almost everyone.
How to start using loops today (the 2-minute version)
You don't need code. You don't need to be technical. You can run a basic loop right now, inside ChatGPT or Claude, with one message. Paste something like this:
That's a loop. You gave it a goal, a way to judge "good," and a stop point. Now it does the back-and-forth for you instead of you doing it by hand.
Try it on captions, emails, cold messages, product descriptions, anything you'd normally nitpick five times. Once you feel the difference, you won't go back to one-shot prompting.
Level up: the real loop setup (loop engineering)
The in-chat version is the gateway. The full version is where it gets scary good. Google engineer Addy Osmani gave the bigger pattern a name: loop engineering. The idea is to replace yourself as the person who prompts the agent. Inside a tool like Claude Code, you wire up a few building blocks so the loop runs on its own around the clock:
Claude Code even ships commands for this. /goal lets you define the finish line, and a separate model checks "are we done yet" after every round. /loop and /schedule keep it running.
Stack those together and it stops feeling like a chatbot. It starts feeling like you've got employees working while you sleep.
When NOT to use a loop
Real talk, because nobody else says this. A loop is not magic. If you can't clearly define "good," a loop will just burn time and money chasing a finish line that doesn't exist.
If the task is fuzzy or creative and you'd only know it's right when you see it, do that one by hand, or at least define the finish line before you let it run. The loop is only as smart as the check you give it.
Define the finish line first. Then let it run.