Stop prompting Claude.
Start writing loops.
The creator of Claude Code and the creator of OpenClaw said the same thing within weeks of each other: you shouldn't be prompting agents one message at a time — you should be designing loops that prompt them. Here's the full playbook, and the three guardrails that keep a loop from running confidently wrong all night.
Get the full loop playbook
The Loop Ladder · the 3 guardrails · loop primitives reference · my exact overnight setup · failure modes. Free. Yours immediately.
Most people use coding agents like a chat box — type a message, wait, type the next one. That makes you the bottleneck. Every minute the agent isn't working is a minute you're typing.
Loops flip it. You write the instruction once; the loop re-issues it — on an interval, until a goal is met, or as an orchestrator spawning other agents. Boris Cherny shipped 259 pull requests in 30 days this way, every one written by Claude. The skill stops being prompting and becomes loop design.
"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."— Peter Steinberger (creator of OpenClaw), June 2026 · 6.5M views