Cherny predicts software developers will broadly transition to this workflow
"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. And this is transition we're going to see for the rest of the year."
- Boris Cherny, head of Claude Code at Anthropic.
Positive users praise the shift to writing loops for managing Claude as a more powerful future approach than prompts, while negative users dismiss the idea over high token costs and code quality risks.
Head of Claude Code Boris Cherny at Anthropic's annual developer conference in San Francisco on how his life changed in the last 6 months with AI writing all the code.
"About six months ago what happened is all the code that I used to have written by hand now Claude writes.
And I just prompt Claude. So I talk to Claude and I'm like, hey, let's build this feature. It builds a feature and it tests it. And then it shows me.
And I'm like, yeah, that's good. Or no, wait, make this change. And then it makes the change. " ---
From 'CNBC Television' YT channel (link in comment)
"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. And this is transition we're going to see for the rest of the year."
- Boris Cherny, head of Claude Code at Anthropic.
Watch full episode at WorkOS YouTube channel:
https://www.youtube.com/watch?v=RkQQ7WEor7w&t=1s
unlimited token on tap and no linkedin is the new FIRE
"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. And this is transition we're going to see for the rest of the year."
- Boris Cherny, head of Claude Code at Anthropic.
This is such a different way of thinking about AI tools. 🤖
Most people are still in the "type a prompt and wait" mindset but the real power is in building systems that run themselves. When the person who built Claude Code is using loops instead of prompts, that tells you everything about where AI workflows are heading. Are you still prompting manually or have you started building loops? 👇
Head of Claude Code Boris Cherny at Anthropic's annual developer conference in San Francisco on how his life changed in the last 6 months with AI writing all the code.
"About six months ago what happened is all the code that I used to have written by hand now Claude writes.
And I just prompt Claude. So I talk to Claude and I'm like, hey, let's build this feature. It builds a feature and it tests it. And then it shows me.
And I'm like, yeah, that's good. Or no, wait, make this change. And then it makes the change. " ---
From 'CNBC Television' YT channel (link in comment)
@rohanpaul_ai Agent loops are basically just state machines with fuzzy transitions. The hard part isn't the loop itself, it's the error boundaries that keep it from burning $200 on an infinite retry bug.
AI works better when you give tools and freedom instead of forcing them into rigid, hand-designed workflows—because general learning systems scale better.
"Ask not what the model can do for you, ask what.."
~ Claude Code creator Boris Cherny (@bcherny)
"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. And this is transition we're going to see for the rest of the year."
- Boris Cherny, head of Claude Code at Anthropic.
https://www.youtube.com/watch?v=kRgdkOw82F0
Head of Claude Code Boris Cherny at Anthropic's annual developer conference in San Francisco on how his life changed in the last 6 months with AI writing all the code.
"About six months ago what happened is all the code that I used to have written by hand now Claude writes.
And I just prompt Claude. So I talk to Claude and I'm like, hey, let's build this feature. It builds a feature and it tests it. And then it shows me.
And I'm like, yeah, that's good. Or no, wait, make this change. And then it makes the change. " ---
From 'CNBC Television' YT channel (link in comment)
@rohanpaul_ai All I ever hear him talk about is how he has his claude set up and he writes no code but what is he having claude actually build is never mentioned
@rohanpaul_ai “I ride jet skis all day everyday. I think we’re going to see a lot of that this year” - CEO of a jet ski company
@rohanpaul_ai @chollet said the same thing from the opposite direction: code volume does not represent productivity. the workflow changed and the metric broke at the same time
@AnatoliKopadze i want the whole world to see this post!
this could change a lot of things
@rohanpaul_ai Writing loops that prompt the model is just prompt engineering with extra steps. The real skill is knowing which loops to kill when they start reinforcing their own assumptions.
@AnatoliKopadze Secure information on this Video
@rohanpaul_ai I have a £19k a month plan. I can barely ask three questions
@rohanpaul_ai Another example of Anthropic being culturally blind when it comes to cost
The better way is a bounded execution system:
- Workflow graph: explicit steps, dependencies, retries, and stop conditions. - Typed I/O: every agent call has a schema, not just text in/text out. - Evaluator gates: loops advance only when checks pass, not because the model says it is done. - Trace and replay: every prompt, tool call, artifact, error, and decision is inspectable. - Supervision layer: a coordinator decides what should run next; the LLM does not silently mutate its own mission forever. - Memory boundary: retrieval and persistence are separate from execution, so the loop is not also the project manager, database, and runtime.
This was the target workflow I had in mind when I wrote llmff. I also used it when building my recursive agent skill improvement sidecar, tugboat (https://syndicalt.github.com/tugboat).
https://www.github.com/syndicalt/llmff
@rohanpaul_ai And guess what, Claude Code is not a great piece of software. The fact that it was loop-agent-generated did not make it exceptional in any way. It didn’t make it “far ahead” of what could have been written by hand.
@rohanpaul_ai Well this is all bloody lovely thinking but not very many people could afford the API cost of running such experiments - THOUGH I am sorely tempted to spin up some open source model and have a crack at it
@rohanpaul_ai yeah yeah yeah, you have unlimited tokens, we hear you
Cherny predicts software developers will broadly transition to this workflow
"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. And this is transition we're going to see for the rest of the year."
- Boris Cherny, head of Claude Code at Anthropic.