Stop Prompting Claude — Loop Engineering and the Four Things Every Loop Needs

Claude Code creator Boris Cherny said the quiet part out loud: 'I don't prompt Claude anymore. My job is to write loops.' Peter Steinberg, creator of Open Claude, posted the same idea on Twitter — stop prompting coding agents, design the loops that prompt them instead.

AI Twitter is full of hype, but when two of the people who built the tools we use every day land on the same point on the same week, it earns attention. The video author takes the claim seriously and breaks loop engineering into three parts: what a loop actually is, the four ingredients of a successful loop, and how a non-technical person can build their first one today.

A normal prompt runs once and stops. A loop is a prompt structure that runs over and over until a specific goal is complete. Boris's own framing: loops are now the things prompting Claude, and his job moved one level up the abstraction stack.

For a loop to work, it needs a goal and a way to verify the goal. The two are inseparable — without verification, the loop has no stopping condition. Technical tasks make this easy. 'Deploy this site and make sure it loads in under 2 seconds' has obvious verification: hit the domain, check content, time the load, and run /engineer-review for the final yes/no.

Non-technical loops are about bridging an abstract goal to a verifiable signal. The /engineer-review skill collapses 'is the code good?' into a yes/no answer. The same move works on fuzzier tasks — find a way to verify the final result even when it isn't strictly measurable.

The operating rule: the less quantifiable your goal, the smaller the sub-goals you should break it into, with checkpoints at each one. Treat the AI like an intern. 'Plan a corporate party' could go anywhere; you'd still want to make the date, venue, and theme calls yourself. Same with loops — mark the forks where one wrong turn ruins the rest of the run, and verify those by hand. The author's closing advice: pick something small and build a single loop orchestration skill for it. That's how the muscle gets built.