DEV Community

Yurukusa
Yurukusa

Posted on • Edited on

4 Hooks That Let Claude Code Run Autonomously (With Zero Babysitting)

#bash #opensource #claudecode #automation

I built a 15,000-line game with Claude Code while sleeping. Here are the 4 hooks that made it possible.

The Problem

Claude Code is powerful, but it stops every few minutes to ask "should I continue?" or "which approach?" In unattended sessions, that means your AI sits idle until you come back.

The Solution

Claude Code Ops Starter — 4 production-tested bash hooks that enforce autonomous decision-making.

What's Inside

1. Context Monitor (context-monitor.sh)

Counts tool calls as a proxy for context window usage. Warns at 3 thresholds:

  • Soft (80): "Consider deferring large tasks"
  • Hard (120): "Wrap up and hand off"
  • Critical (150): Auto-generates checkpoint file for session handoff 
# Configurable thresholds
SOFT_WARNING=80
HARD_WARNING=120
CRITICAL=150
Enter fullscreen mode Exit fullscreen mode

2. No-Ask-Human (no-ask-human.sh)

Blocks all AskUserQuestion tool calls. Instead of stopping to ask, the AI:

  1. Decides on its own
  2. Logs uncertainty to ~/pending_for_human.md
  3. Moves to the next task

Override with CC_ALLOW_QUESTIONS=1 when you want questions back.

3. Syntax Check (syntax-check.sh)

Auto-runs verification after every file edit:

  • Python: python -m py_compile
  • Shell: bash -n
  • JSON: jq empty

Catches errors immediately instead of 50 tool calls later. Gracefully skips if jq or python aren't installed.

4. Decision Warn (decision-warn.sh)

Flags rm -rf, git reset --hard, DROP TABLE before execution. Warns but doesn't block — uncomment one line for hard blocking.

Install 

git clone https://github.com/yurukusa/claude-code-ops-starter.git
cd claude-code-ops-starter
bash install.sh
Enter fullscreen mode Exit fullscreen mode

No dependencies besides bash. jq and python are optional (hooks skip gracefully without them).

How It Works

Claude Code has a hooks system that runs shell scripts at specific lifecycle points:

  • PreToolUse: Runs before a tool call (can block it with exit 1)
  • PostToolUse: Runs after a tool call (for monitoring and validation)

The starter kit uses both:

Hook Trigger Effect
no-ask-human.sh PreToolUse (AskUserQuestion) Blocks with exit 1
decision-warn.sh PreToolUse (Bash) Warns (exit 0)
syntax-check.sh PostToolUse (Edit/Write) Validates (exit 0)
context-monitor.sh PostToolUse (all) Monitors (exit 0)

Background

These hooks come from 200+ hours of real autonomous Claude Code operation. Every hook solved a real problem:

  • The context monitor was born from sessions that ran out of context without warning
  • No-ask-human from overnight sessions where the AI sat idle for hours waiting for a human answer
  • Syntax check from discovering errors 50 tool calls after they were introduced
  • Decision warn from a git reset --hard that wiped an afternoon of work

These 4 hooks are the foundation. The full system is the house. 15 minutes to set up what took us 200 hours to build — the CC-Codex Ops Kit includes multi-agent orchestration, stall detection, watchdog, task queue, and 20+ hooks.

GitHub | @yurukusa_dev on X

Before your next overnight run: Scan your Claude Code setup for free — 30 seconds, nothing installed. See if your 4 hooks cover the critical gaps.

Free Tools for Claude Code Operators

Tool What it does
cc-health-check 20-check setup diagnostic (CLI + web)
cc-session-stats Usage analytics from session data
cc-audit-log Human-readable audit trail
cc-cost-check Cost per commit calculator

Interactive: Are You Ready for an AI Agent? — 10-question readiness quiz | 50 Days of AI — the raw data

Top comments (0)