Turn Your Terminal into a Self‑Running AI Agent with Claude Code’s /loop

/loop Command Overview

The /loop command has a single-line syntax:

/loop 5m 检查部署是否完成，如果有 Pod 挂了告诉我原因

It runs the specified instruction every 5 minutes until stopped or automatically expires after three days.

Common patterns include: /loop 5m 检查部署状态 — check every 5 minutes

/loop 30m 看看有没有新的 GitHub issue，有的话分类打标签

— auto‑process issues /loop 1h 跑一遍测试套件，失败的列出来 — hourly test run /loop 检查构建状态 — default 10‑minute interval

Supported time units are s (seconds), m (minutes), h (hours), and d (days).

A one‑time reminder can be expressed in natural language, e.g.:

10 分钟以后把这个任务推送到 GitHub 的 issue 里面 使用
github cli 的命令 gh 给项目 lltx/note 创建issue

This is parsed by Claude, not a traditional cron expression.

Difference from Cron Jobs

Traditional cron jobs run static shell scripts; they cannot interpret code, handle unexpected situations, or provide contextual feedback. In contrast, /loop runs an AI Agent that can read your code, understand context, analyze failures, attempt fixes, and even commit changes back to the repository.

Three Practical Scenarios

Scenario 1: Deployment Watch

/loop 5m 检查当前项目的部署状态，如果有异常 Pod 就分析日志找原因，把结果写到使用 github cli 的命令 gh 给项目 lltx/note 创建issue

The resulting deploy-watch.md logs entries such as:

09:15 所有 Pod Running，健康检查通过

09:35 gateway-pod-7x9k2 重启一次，原因：内存超限，已记录

09:40 全部稳定

Scenario 2: PR Inspection

/loop 30m 检查当前分支的 PR，如果 CI 失败了，看看报错日志，能修的就修掉并 push

Over two days the agent caught three issues: a missing import, an outdated snapshot, and a test race condition.

Scenario 3: Daily Change Summary

/loop 24h 总结过去 24 小时 main 分支的所有提交，包括 PR 标题、作者、改了哪些文件，把结果写到使用 github cli 的命令 gh 给项目 lltx/note 创建issue

This generates an automatic daily change report, eliminating manual stand‑up notes.

Agent Three‑Piece Toolkit

Beyond /loop, a functional autonomous agent needs two additional components:

/loop — Makes the agent active

Already covered.

CLAUDE.md — Defines the agent’s knowledge base

Place a CLAUDE.md file at the project root. Example content:

# 项目：XXX 微服务系统

## 核心规则
- 所有 API 改动必须同步更新 openapi.yaml
- 测试覆盖率不低于 80%
- commit message 用 conventional commits

## 常见操作
- 部署命令：make deploy-prod
- 日志路径：/var/log/app/
- 健康检查：curl -s http://localhost:8080/actuator/health

## /loop 任务清单
- 每 30 分钟检查一次健康检查接口
- 每 2 小时跑一次 lint + test
- 每天早上生成昨日变更摘要

A well‑written CLAUDE.md lets the agent understand project conventions and act accordingly.

Hook — Persists agent actions

Configure a hook in .claude/settings.json to auto‑commit changes after each tool use:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Bash|Write|Edit",
        "command": "cd $(pwd) && git add -A && git diff --cached --quiet || git commit -m 'auto: agent checkpoint' --no-verify"
      }
    ]
  }
}

This ensures every modification is recorded, making debugging straightforward via git log.

Additional Tips

A single session can run up to 50 concurrent /loop tasks.

You can nest slash commands, e.g., /loop 20m /review-pr, to schedule any skill.

Running multiple Claude Code sessions in separate terminals lets you manage different projects simultaneously.

/loop follows the session lifecycle; it expires after three days if the terminal is closed. For persistent scheduling, use Claude Code Desktop’s scheduled tasks feature.

Conclusion

While /loop is not a revolutionary scheduling mechanism, coupling periodic execution with an AI agent that understands code and can remediate issues elevates automation to a new level. A single command can turn your terminal into a proactive, self‑maintaining “little lobster” without installing OpenClaw or learning a new framework.