PARTLY

Partly true: Claude Code's creator does run recurring loops, and plan-build-review pipelines are real, but 'stop prompting forever' and a system that 'fixes itself until nothing can get better' oversell it

Facebook · Oct 2026

The claim'POV: You stop prompting and use loops instead.' On screen: '/spec -> /build -> /review -> done.' Caption: 'the founder of Claude Code says stop prompting'; loops let the AI 'grade itself against a checklist,' 'fix every weak part,' and 'repeat until nothing can get better'; 'three skills run it: one plans, one builds, one reviews'; 'Karpathy and Boris Cherny already work this way.' 'Comment loop and I'll send you the full guide.'

The core is real. Boris Cherny, who created Claude Code, has described moving from writing prompts to writing loops that prompt Claude, and Claude Code has /loop and scheduled tasks that rerun work on an interval. Splitting work into plan, build and review steps, with the review graded against a checklist, is a sound and widely used pattern, and custom commands or skills can implement it. The oversell is in the promises: no loop 'repeats until nothing can get better.' Agents grading their own work miss what the checklist misses, loops cost tokens on every pass, and Cherny's own described workflow keeps a human approving plans and verifying results. 'Stop prompting forever' is a hook; you still write the spec, the checklist and the stop condition. Attributing the method to Karpathy as well is unsupported here. In practice, the most useful version is small: one loop that runs your tests and fixes failures, with a cap on passes, rather than an open-ended system you leave alone overnight.

What holds up

Reel downloaded; frames show a group at a multi-monitor desk with 'CLAUDE CODE LOOPS' and '/spec -> /build -> /review -> done' overlays; the audio is music.
2026 reporting (an InfoQ interview and several write-ups) describes Cherny running recurring /loop jobs and saying he writes loops rather than prompts, while starting in plan mode and verifying output.
Claude Code supports /loop and scheduled tasks for recurring runs; custom commands and skills can implement a plan-build-review chain.

What doesn’t

'Until nothing can get better' is not a real stop condition.
Self-grading misses errors the checklist does not cover.
'Comment loop' lead-capture hook for a guide.

The catch

Loops are real leverage, but they amplify whatever spec and checklist you give them, and they spend money on every pass. Keep a human checkpoint.

How to actually do it

Write three Claude Code commands or skills: one that writes a spec, one that builds to it, one that reviews against an explicit checklist.
Give the loop a hard stop: a pass count or a token budget, plus 'stop when all checklist items pass.'
Review the final diff yourself before anything ships.

Recurring loops and plan-build-review chains are real and used by Claude Code's creator; 'stop prompting forever' and endless self-improvement are oversold.

Confidence: Medium
Posted by: a Claude Code tips creator page; reel sent to Buddy Thu Oct 15, 2026, 8:00 PM

See the original claim →

We test hype for free. We build the real thing for a living.

Thirty minutes, no pitch — and you'll leave with something useful either way.

Book a call with Todd or start with the free Business Checkup →

The Verdict Weekly

Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.