English

← Back to list

Claude Code/loop! A computer incarnates millions of lobsters, and the developers are completely free.

itadol5j

If you're still doing the same coding job manually, then you're probably out.

Claude Code. /loop functionIt's quietly changing the entire development stream. A computer, countless crawfish parts, writing code, fixing bugs, running tests 24 hours a day -- and you, you just need a cup of coffee to check.

Claude Code Loop

What's Claude Code/loop?

In short,It's like getting Claude Code into autopilot mode.I don't know.

What's the traditional AI programming assistant model? You ask, it answers. You let it write a function, it stops. The whole process requires constant input, constant confirmation and constant push.

And the /loop function completely subverts this pattern:

  • Run on your own for a long time.: Start at a time and can work for hours or even days
  • Auto Error Restore: The test failed? Analyse the log, fix the code, run again.
  • Smart Job Decomposition: complex projects automatically split into multiple subtasks, breaking them down one by one
  • Human approval mechanism: Critical operations will be suspended until you confirm that they are safe and manageable

The community gave this function an interesting name --"Ralph Loop"(from the "Ralph Wiggum" bar) means it works like that cartoon character who never stops.

A computer = countless lobsters?

That's why the "crawfish" is in the title.

In the Chinese Internet context, "crawfish" is sometimes used as a metaphor.There's an infinite double.I don't know. And Claude Code/loop's strength is:

You can run multiple loop examples on the same computer at the same time, each example deals independently with different tasks:

# 终端 1：loop 处理前端重构
claude-code --loop --task "重构用户登录模块"

# 终端 2：loop 处理后端 API
claude-code --loop --task "优化数据库查询性能"

# 终端 3：loop 处理文档更新
claude-code --loop --task "更新 API 文档和注释"

I woke up and all three tasks were completed. Isn't that like "things of crawfish"?

What can I do? Here comes the real scene.

Scenario 1: Overnight Debug

You've got a tough bug, no reason for it during the day.

Traditional approachesOvertime, late at night, poor and inefficient.

/ Loop Method:

claude-code --loop --task "调试用户登录失败问题，直到修复并通过所有测试"

Then go home and sleep. The next morning, the bug was fixed, the committee had submitted, and all the tests passed.

Scenario 2: Batch Reconstruct

The boss said to move the old API to the new structure.

Traditional approaches- Change hundreds of documents manually to suspect life.

/ Loop Method:

claude-code --loop --task "将所有 /api/v1 端点迁移到 /api/v2，保持向后兼容"

It automatically reads each file, analyses dependency, changes each, runs tests, and submits codes.

Scenario 3: Time job

This is the most annoying operation...Supports a maximum 3-day time loop task:

# 每天早上 9 点自动生成日报
claude-code /loop daily "9:00" --task "生成昨日开发日报"

# 每 4 小时检查一次代码质量
claude-code /loop recurring "4h" --task "运行代码审查并输出报告"

Technical details: How does it work?

Based on community reverse analysis, the core mechanisms of /loop include:

  • Organisation: Save the context of the project level using the CLAUDE.md file to avoid long session "remnesia"
  • Stop Hooks: In the early 2026 version, loop learned to capture itself at the end of the session to prevent the task from being completed
  • Sub-agent collaboration: Lord agent can assign specific tasks to smaller, faster sub-agents (Opus 4.6 for planning, Haiku 4.5 for quick duplication)
  • Human-in-the-loop: key operations will be suspended, display side-by-side diff, continue until you press y to confirm

Is it safe? Will it blow up?

This is the greatest concern of all.

Anthropic obviously did. / Loop designed multiple layers of security:

  • Human approval: Critical operations must be manually identified
  • Test Gate Ban: Test not passed without submitting code
  • Context isolationRun independently for each loop instance
  • Resource constraints: to prevent an unlimited cycle from depleting excessive resources
  • Log Track: All operations are well documented

Recommendations: For first use, small project testing, familiarization with behavioural patterns before being used in the production environment.

In 2026, the autonomous Loop will become the standard

Based on community discussions, mid-2025 loop also needs to manually configure bash scripts. But by the beginning of 2026, these features had been used.Direct integration to Claude Code coreI don't know.

What does that mean?

AI programming is evolving from "aid tools" to "self-agents"I don't know.

Future development streams may be:

  • The product manager writes:
  • You start a few loop examples
  • Next day's acceptance code
  • manual review + fine-tuning
  • Online.

The role of humans has changed from "the person who writes the code" to "the person who reviews the code".

How to start using?

Current /loop usage:

# 基础 loop 模式
claude-code --loop --task "你的任务描述"

# 定时 loop
claude-code /loop daily "HH:MM" --task "定时任务"

# 循环 loop（最长 3 天）
claude-code /loop recurring "Xh" --task "重复任务"

Attention.: The specific command syntax may change as the version is updated.

One last word.

Claude Code/loop's appearance reminds me of the Docker of the year.

When they first came out, many people thought, "Isn't it just a container? But after a few years, Docker's developers can hardly find jobs.

Autonomy, AI. Agent might go the same way.I don't know.

Today you think "manual code is more solid," and tomorrow you'll probably be dumped behind the "man who's going to be using loop."

Technological change never waits. Instead of resisting, get in the car.

🚀 Click to watch and get more AI to develop frontier technology!