Stop prompting. Build your first AI loop.
Boris Cherny built Claude Code. In an interview in June he said the quiet part out loud: “I don't prompt Claude anymore.” He has loops running that prompt it for him.
That sounds like a flex until you notice what it actually says about the rest of us. He did not stop using AI. He stopped being the part of the process that reads the answer, spots the mistake, and asks again.
What a loop actually is
Strip away the jargon and a loop has three ingredients. Miss any one of them and you do not have a loop.
- A goal you state once. Not a chat. One instruction, written properly, at the start.
- A check the machine can run without you. Something that comes back pass or fail on its own. This is the ingredient people skip, and it is the one that matters.
- A stop condition. The check passes, or the budget runs out. Otherwise it spins forever.
Put together, the AI plans, builds, checks its own work, and goes again when the check fails. Anthropic documents this shape as the evaluator-optimizer workflow: one pass produces the work, another grades it, and the grade feeds back in.
The check is the whole trick
Cherny's single biggest tip for getting good results is not a clever prompt. It is giving the AI a way to verify its own work. He puts the difference at two to three times the quality of the final result:
A real check has one property: it can fail. Asking the AI “does this look right?” is not a check, because the thing that wrote the mistake is the thing grading it. These are checks:
- A test suite, or a single test that must pass.
- A build or compile command that either succeeds or errors.
- A script that opens the page and reads back the number that should be there.
- A second AI with the original goal in hand and one job: say pass or fail, and why.
Your first loop, in about twenty minutes
Do this with whatever agent tool you already use. The prompts below are copy-paste. The only real work is picking the job in step one.
Pick a job that can fail out loud
Good first loops: make this test pass, get this script to run clean, keep reformatting this file until the linter is silent, fix every broken link on this page. Bad first loop: make my landing page better. Better has no fail state, so the loop never ends and nothing improves.
State the goal once, properly
You get one shot at explaining it, so include what done looks like. This is the part you would normally dribble out over fifteen messages.
Your goal: [the outcome, in one sentence]. Done means: [the exact condition someone else could verify]. Write a short plan first, then start. Do not ask me questions along the way. If something is ambiguous, choose the simplest option and note the choice.
Hand it the check, as a command
Name the check explicitly and make the AI run it, not you. If it has to ask you whether it worked, the loop is broken at the only joint that matters.
After every change, run this exact command and read the output: [your command]. Treat its output as the only source of truth about whether you are done. Do not tell me it works. Show me the output of the command.
Check: break something on purpose and run the loop. If it reports success, your check is not really checking.
Set the stop condition and the budget
Two ways out: the check passes, or the loop gives up and tells you what it learned. Both must exist before you walk away.
Repeat this cycle until the command passes, or until you have tried 5 times, whichever comes first. If you hit 5 attempts, stop and report what you tried, what the output was each time, and what you think the real blocker is.
Make it remember, so it stops repeating itself
A loop that forgets makes the same mistake all night. Have it write what it learned into a notes file the next run reads first. In Claude Code that file is CLAUDE.md, but any file the tool reads on startup works.
Whenever a fix fails, append one line to NOTES.md: what you tried and why it did not work. Read NOTES.md before your next attempt and do not repeat anything on that list.
When not to loop
Loops are for work with a cheap, honest check. Two categories stay out. Taste work, where the only judge is a person, and anything you cannot take back: paying, sending, publishing, deleting. The move there is to let the loop do everything up to the last step, then stop and wait for a human to press the button.
If you cannot name the check, you do not have a loop, you have a wish. Name the check first and the rest of it is plumbing.
None of this is fringe. When Peter Steinberger, who created OpenClaw, posted the same advice in June, that one-line version of it collected more than six million views in a week, and the argument about it ran for days. The idea is not the hard part. Writing the check is.
Questions people actually ask
Do I need to know how to code to run a loop?
No, but you do need a check the machine can run by itself. If your check is a command, a test, or a script that either passes or fails, an AI coding tool can run the loop for you. If your only check is your own opinion, you do not have a loop yet, you have a conversation.
How is this different from just asking for a plan?
A plan is one answer. A loop is a plan plus a check plus permission to try again. The AI runs, the check fails, the failure goes back in, and it runs again without you reading anything. You only look at the end.
Will it run forever and cost a fortune?
Only if you forget the stop condition, which is why it is one of the three ingredients here. Give every loop a budget in attempts or minutes, and make it stop and report rather than keep grinding. Set the budget before you set it running, not after.
What kind of work should never go in a loop?
Anything with no cheap check and anything you cannot undo. Sending email, paying, deleting, posting publicly. Let the loop prepare the work and stop at the last step so a human presses the button.
Sources
- Boris Cherny (creator of Claude Code) on verification being worth 2-3x the final quality
- Boris Cherny on self-verification loops letting a model run far longer
- WorkOS — takeaways from the Acquired interview where Cherny describes writing loops instead of prompts
- Peter Steinberger (creator of OpenClaw) — design loops that prompt your agents
- Anthropic — Building effective agents (the evaluator-optimizer workflow)
- Anthropic — Claude Code best practices
Want this built for you?
We write these memos because we build this stuff every day. If you want it working in your business instead of sitting on your reading list, that is literally our job.