A tweet claiming an "Anthropic engineer" said you're not supposed to prompt Claude, you're supposed to build a system that prompts itself has pulled 1.2 million views. The pitch is simple: stop writing one prompt at a time, start designing loops that keep working after you close the laptop.
Here's my read after going through it. The technique is correct and most builders aren't using it. The story wrapped around the technique is engagement bait, and the accounts promoting it can't agree on basic facts.
The part worth keeping
Four claims sit at the center of the thread: you skip the memory file so every loop starts from zero, you never split your sub-agents, you have no stop condition so loops run forever and bill you in your sleep, and writing one prompt a day is the slowest way to use Claude.
Each of those maps to a real Claude Code capability. The memory file is CLAUDE.md, which persists context across sessions so a loop doesn't re-learn your project every iteration. Sub-agents let you hand a sub-task to a separate context instead of cramming everything into one thread. The stop condition is what keeps an autonomous loop from spinning indefinitely on your dime. None of this is exotic, and that's the point: the gap isn't capability, it's that people are still treating Claude like a chat box.
So the central argument holds. A prompt gets you one response; a Claude loop gets you a system that keeps producing after you walk away. If you've used Claude for a month and still type every task by hand, you are leaving the most valuable mode of the tool switched off.
What builders actually said
The reception was loud and almost entirely positive. One account called it a 28-minute video and said "I've seen $300 courses that don't cover what he shows in the first 10 minutes," highlighting CLAUDE.md files, memory shortcuts, and parallel sessions as the standouts. [1]
“The engineer who built Claude Code just dropped a 28-minute video on how to write prompts that actually work I've seen $300 courses that don't cover what he shows in the first 10 minutes CLAUDE.md files, memory shortcuts, parallel sessions, prompting patterns all in one video and completely free works whether you're a developer, a beginner, or someone who's been using Claude for months.”
The sharpest framing came from a different angle. One builder argued the real lesson isn't prompting at all: "a great prompt still dies the second you close the tab," which is the cleanest one-line case for loops I've seen in the whole thread. [2] He also leaned hard on credibility, calling it Anthropic's own applied AI team, "the people who wrote the weights, showing how they do it." [2]
“Anthropic's own applied AI team just dropped a 24-minute workshop on how to actually prompt Claude - free not a creator who reverse-engineered it. not a Reddit thread. the people who wrote the weights, showing how they do it. and the quiet takeaway is bigger than prompting: a great prompt still dies the second you close the tab watch what they're really teaching: > structure, examples, and giving ”
That credibility claim is exactly where I get cautious.
Where the story falls apart
The sources can't agree on who made this or what it is. The original thread attributes the loops philosophy to Boris Cherny, who runs Claude Code at Anthropic. One promoter calls it "the engineer who built Claude Code" and a 28-minute video. [1] Another calls it "Anthropic's own applied AI team" in a 24-minute workshop. [2] An individual engineer, a team, 28 minutes, 24 minutes. Those are not the same claim, and the discrepancy runs through the very tweets selling the video as authoritative.
Then there's the math. The thread asserts that 95% of builders have never split their sub-agents. No source backs that number, and it carries the unmistakable shape of a stat invented to make you feel behind. Treat it as rhetoric, not data.
I want to be precise about what I'm criticizing. The loop technique is sound. The packaging, the bookmark-before-it-gets-buried urgency, the conflicting attributions, the round-number statistic, is the standard machinery for making a free video feel like a secret. You can take the substance without buying the theater.
Where this leaves us
Build the loop. Skip the hype.
Concretely, start with the memory file. A CLAUDE.md that captures your project's structure and conventions is the highest-leverage change here, because it's what makes every subsequent loop iteration smarter instead of amnesiac. Add sub-agents when a task has a clearly separable piece worth isolating, not before. And set the stop condition first, not last: an autonomous Claude loop with no exit criteria is the one failure mode that costs real money while you sleep.
On the video itself, watch it if you want the walkthrough, it's free, but verify the attribution before you repeat any "Anthropic says" claim built on top of it. The technique stands on its own merits. It doesn't need a contested origin story to be worth your evening.
Sources
- Tweet by @keitijeon — @keitijeon · community
- Tweet by @kingwilliam_ — @kingwilliam_ · community
- Tweet by @mikenevermiss — @mikenevermiss · community