From prompt crafting to loop engineering

Stop prompting, instead let AI prompt itself

Apparently our job prompting is done here.

Our job moving forward is to write loops, says Boris Cherny who leads Claude Code.

Let AI prompt itself

I’m asking you in the post to hear Cherny out (and to consider the alternative.)

Design systems where AI talks to itself.

The head of Claude Code at Anthropic says he doesn’t prompt Claude anymore.

Boris Cherny just said this publicly:

“I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.”

Become a loop architect

Read that again. The person building Anthropic’s coding agent doesn’t write prompts.

He writes loops that let the AI prompt itself.

And says this is the transition we’ll see for the rest of the year.

The skill isn’t talking to AI. That is what we have been doing when we manually prompt into ChatGPT or Claude.

The skill is designing systems where AI talks to itself.

Cherny advocates for building autonomous self-correcting workflows instead of manually feeding AI individual instructions.

Loop engineering is about setting up a recursive, continuous control system.

Photo: Craig Blankenhorn/Max

Picture one of the most iconic scenes from Sex in the City’s Carrie Bradshaw sitting on her bed, typing into her Apple laptop:

Are most developers building for a world that’s already gone?

The image of her sitting among the pillows, ruminating in her Upper East Side apartment, and typing her famous column intros has become a defining pop culture trope.

We don’t need to understand the evolution of Carrie’s Apple laptops even though throughout the original HBO series and subsequent films, Carrie’s choice of technology perfectly matched her fashion-forward sensibilities.

It is not important for us to understand that in the PowerBook Era, for much of the original series, Carrie typed away on an original Apple PowerBook G3, which led to the “Sad Mac” Incident, where her fraught relationship with tech is highlighted in Season 4, Episode 9 (”The Agony and the ‘Ex’-tacy”). After Aidan kisses her while she’s typing, she crashes her PowerBook and gets the dreaded “Sad Mac” screen. Her total lack of backing up her work results in a frantic trip to a repair shop. More recently, in the HBO Max reboot And Just Like That... Carrie upgrades her setup slightly but continues to cling to her beloved, older-model MacBook Pro. Today, Carrie could just let AI prompt itself.

It isn’t even important for us to recognize that Vogue published a piece this time last year, An Ode to Carrie Bradshaw’s Laptop, My Favorite Sex and the City Character

The idea that the laptop is a favorite character of the TV show isn’t worth considering. AI has none of the appeal of an Apple laptop.

Carrie’s famous signature internal questions

This is what is important. When Carrie sat down and typed

she was formulating questions that led not only to her column but became the central question of each episode.

What is important for the sake of this post are not her famous signature internal questions, but how she struggled to arrive at them.

Carrie Bradshaw’s column introductions kicked off with one of her famous signature internal questions. While typing on her bed with a cigarette in hand, she often pondered the dating habits of New Yorkers, framing the overarching theme of that week’s episode. The specific topics matter less than the fact that she asked them is what is important as we look at the impact on our lives from loop engineering.

Some of Carrie’s most iconic opening column questions include:

In The Age of Uninnocence episode, she typed:

“Welcome to the age of un-innocence. No one has Breakfast at Tiffany’s, and no one has Affairs to Remember.”

In the Monogamy in the City episode:

“In a city like New York, with its infinite possibilities, has monogamy become too much to expect?”

In The Rules of the Game episode:

“After all, computers crash, people die, relationships fall apart. The best we can do is breathe and reboot.”

And poignantly for our purposes here, in the Past vs. Future episode, Carrie asks:

“I couldn’t help but wonder... can you get to a future if your past is present?”

AI will never work for people until we get it to sound like Carrie Bradshaw typing into her laptop on her bed while talking to herself in her head.

That’s us. Right now. Thinking up what to prompt. Trying to articulate what it is we want to say.

We are in other words homo sapiens interrogans, the "question-asking animal.”

The act of formulating a question is cognitive work. When Carrie sat down and typed

she was converting a messy lived experience into a focused inquiry. That translation process, from raw feeling or observation to a precise question is where a lot of human meaning-making happens. It’s the moment you discover not only what you think but, for me at least, that I think.

Loop engineering hands that moment off. You give Claude an intention, and it generates the question, refines it, responds to itself, and eventually surfaces an answer. The human becomes the approver of a process rather than the author of an inquiry. There’s a loss there, and I’m arguing not just a philosophical one. Socrates didn’t think the answers were the important part, the examined asking was.

Here, it’s also worth mentioning investment. Questions you’ve wrestled into existence carry emotional weight. They’re yours. Answers just feel different when you arrived at them through your own questioning and prompting. Let’s not delude ourselves: prompting alone isn’t some escape hatch or reason to think we are exceptional and worthy of taking credit for the answers they lead to. But at the same time, the process that leads to the question we mean to ask is important. A loop-engineered output, however polished, may feel more like something that happened to you than something you discovered.

Where I may be wrong

Or at least where it’s more complicated because loop engineering is, at this writing, just 4-5 days old.

The Carrie Bradshaw image on the bed dreaming up questions then typing is romantic, but it’s worth remembering she was a professional writer. Most people, most of the time, are terrible at formulating the question they actually mean. They ask the surface question, not the real one. A well-designed loop that pushes back and reframes might surface a better question than the person started with. If this is true, then loop engineering doesn’t replace human inquiry but serves to excavate it.

Boris Cherny is also working in a domain, software engineering, where the goal is not self-knowledge but a working artifact. The question “how do I build this?” is instrumental.

Carrie’s question “why do we sabotage love?” is about forming an identity and understanding.

Loop engineering may be appropriate for the first question and impoverished for the second. The loss I am worried about may not be universal but domain specific.

Then there is the calculator question that we keep seeing. When search engines arrived, people worried we’d lose the ability to browse and discover serendipitously. When GPS arrived, we worried about losing spatial intuition. Same with the introduction of the calculator. These concerns had merit but humans adapted, and the skills lost (e.g. remembering everyone’s phone number) weren’t the most valuable ones. The question is whether question-formulation is more like map-reading (losable without immense cost) or more like something closer to the core of what makes thought personal and meaning-laden.

There is a class of questions, the ones where the asking is the point, where the fumbling toward language is how you figure out what you believe that loop engineering would hollow out.

The risk isn’t that everyone will stop asking beautiful, exceptional and insightful questions.

It’s that we’ll increasingly ask only the questions that are answerable, and outsource the uncomfortable, half-formed ones that don’t resolve neatly which are often the most important ones we have.

No human in the loop?

I could end the post here, but I’d like to return to the shift Boris describes, from writing prompts to writing loops and the implications it has for mere mortals who would like to stay in the loop.

During Boris Cherny’s talk at the Acquired Unplugged event hosted by WorkOS on June 2, 2026, he detailed how his daily workflow has completely changed from active programming and manual prompt engineering to what developers are calling loop engineering.

Prompts are more than individual instructions. They come a thoughtful human who through the friction of inquiry often result in surprising responses. What loop engineering makes clear is that the real engineering challenge is designing the feedback structure around the model, not the individual instructions aka prompts into it.

The exact full quote from Cherny’s presentation is:

“I was running 5, 10 Claudes in parallel. My coding was prompting Claude to write Code. Now it’s actually leveled up I think again to the next abstraction where I don’t prompt Claude anymore. I have loops that are running. They’re the ones prompting Claude and kind of figuring out what to do. My job is to write the loops.”

Your job now is to write the loops

What does he mean by writing loops?

Instead of a human sitting on a bed in front of a laptop or at their dining room table in front of a chat box typing out instructions, Cherny writes automated, self-correcting routines.

A script or automated workflow defines a final goal and an evaluation metric (like running a test suite or verification gate).

The program prompts Claude automatically captures the model’s output and tests it.

If the code fails a test, the script feeds the raw error log right back to Claude, commanding it to fix itself.

There is only a human in the loop to the extent that Cherny is needed to write the loops. Afterwards, he apparently goes to bed.

Cherny explained that he regularly spawns a few hundred agents across multiple parallel Claude sessions during the day, letting thousands run deeper tasks overnight while he is asleep.

Unlike Carrie Bradshaw who famously relied on her Apple laptop, Cherny reportedly uses his phone as his primary machine to manage and merge the resulting code.

Prompts vs. loops

Instead of giving the AI a single prompt and reading the output, loop-based automation involves setting up a systematic pipeline.

Where humans currently provide manual input & task initiation, writing loops involves loop design and establishing system boundaries.

Whereas writing prompts is one-and-done, writing loops is continuous, where the loop reads the goal, acts, verifies, corrects, and repeats as necessary.

Instead of assuming the AI is correct, the loop has built-in verification (like unit testing, compiling code, or checking against criteria).

If the step fails, the AI evaluates the error and tries a different approach, looping until the task is complete without human intervention.

When we manually prompt, AI responds to questions. Writing loops operates in the background, autonomously reading and acting on files while we sleep.

Cherny’s philosophy is that the leverage is in building the control system itself rather than fussing over the exact words in a single prompt.

Only fussing over the exact words is as close to a job description I have for who I am and what I do. They’ll have to pry my fussing over words from my cold, dead hands.

The cost of compute

After telling everyone that coding is solved a few months ago, Boris Cherny is trying to assure us that the future is writing loops, that we won’t need to prompt the LLM anymore, just tell it:

“Build X, don’t stop until you are finished” and wait for your immaculate result. It’s as easy as that.”

Only, be sure your loop has a budget cap or it may wipe out your savings.

In the morning, you will wake up to the bill.

Outsource your understand

It’s easy to write loops and leave them running forever if you don’t have to pay for it.

“Someone has to pay for loop. Once you start paying for loop you will understand it’s not wise to outsource your understand.”

We will have to wait and see how it works in practice, how it scales in an enterprise environment, with strict security and compliance requirements.

AI building AI

Writing loops that prompt Claude to prompt itself. Where does that leave us?

The important question is whether the loop knows when to stop because the answer is good enough versus when to stop because it ran out of compute budget.

How, in other words, does the loop know when it is done?

…

This is a developing story. Stay tuned. If you want to get a taste of what using agentic AI on steroids in your sleep will be like, this recent piece by Clive Thompson is excellent and worth reading or listening to.

The Small-Business Owners Managing Whole Armies of A.I. Employees: When you turn A.I. agents loose on your finances, email and customers, what could possibly go wrong?

And this. AI models as Sex and the City characters:

Claude = Charlotte (kind, thoughtful, slightly anxious about doing the right thing)
ChatGPT = Carrie (overshares, occasionally wrong, somehow still the main character)
Gemini = Miranda (smartest in the room, won’t let you forget it)
Grok = Samantha (no filter, says the thing)