Anthropic just dropped a guide on exactly how to use loops in Claude Code.

Anthropic just dropped a guide on exactly how to use loops in Claude Code. Think of a loop as an agent repeating cycles of work until a stop condition is met. And there are 4 types, from beginner to fully hands-off. Type 1 is turn-based. This is what most people do already. You prompt, Claude works, you check the result, you prompt again. Every single prompt is technically a loop... you're just the one running it. Type 2 is goal-based. You type /goal and define what done looks like. The example from the guide. Get the lighthouse score to 90 or above. Stop after 5 tries. Claude can't quit early. Every time it tries to stop, the model checks the condition to send it back to work. Type 3 is time-based. /loop prompt runs a prompt on an interval. Like every 5 minutes, check my PR, address review comments, fix failing CI. The catch is, /loop runs on your computer. Turn it off and it stops. /schedule moves the routine to the cloud, and it keeps running without you. Type 4 is proactive loops. This is where you combine everything. You might schedule a routine to check feedback every hour, then set a goal so it doesn't stop until every bug is triaged and actioned. Use a workflow to explore three fixes in parallel, then have an agent review them. And for this, no human is needed in real time. One warning though, not every task needs a complex loop. Start simple. Level up only when the task demands it. I've summarized the whole guide: all four loop types, the exact prompts, and how to keep token usage in comment. Comment "LOOPS" and I'll send you the link.