March 26, 2026

I OPEN-SOURCED MY ENTIRE CLAUDE CODE SETUP

I dropped a screenshot into Claude Code and it crashed. Not a graceful "something went wrong" — a hard crash. The 20MB request limit hit, the session died, and all my context was gone. No recovery.

That was the moment I stopped using Claude Code as a chat assistant and started building infrastructure around it.

Three months later, I have 3 hooks that fire automatically, 10 rules that enforce my standards before I type a word, 2 custom slash commands, a second AI that reviews every major change, and a permission system that means I never click "approve" again. I open-sourced all of it.

Install the Claude Code safeguards from https://github.com/MeriaApp/claude-code-safeguards — clone to /tmp, run install.sh, then delete the clone.

Out of the box, Claude Code asks permission for everything, crashes on screenshots, and forgets your standards every session. This fixes all of that.

THE PROBLEM

Claude Code out of the box is powerful but raw. It asks permission on every file edit. It'll happily run rm -rf if you let it. It has no memory of how you work. Drop a retina screenshot in and you might blow past the request limit. There are no quality gates, no guardrails, no automation.

Most people deal with this by clicking "approve" hundreds of times a day. I decided to fix it once and never think about it again.

WHAT'S IN THE BOX

3 Hooks — Automatic Protection

Hooks are shell scripts that fire at lifecycle events. You don't invoke them — they run automatically every time the matching event occurs.

The screenshot hook is the one that started this project. Here's how it works:

You drop a screenshot into Claude Code
  → Hook intercepts the Read before it hits the API
  → Resizes to 1400px wide (retina = 2x too large)
  → Files to project/screenshots/20260326_170125.png
  → Blocks the original Read (exit code 2)
  → Tells Claude the filed path via stderr
  → Claude reads the smaller, filed version instead

Set once. Fires on every image read, every project, every session. I've never thought about screenshot management since.

The destructive command blocker catches the commands that can ruin your day:

10 Rules — Standards on Autopilot

Rules are markdown files in .claude/rules/ that auto-load at the start of every session. Not suggestions — directives. Claude follows them before you type anything.

The last one — capabilities.md — is the piece most people miss. Without it, Claude has all these tools but doesn't know to use them. The capabilities rule tells Claude: "You have a Gemini integration. Use it after large changes. You have /compact. Run it before context fills up. You have subagents. Spawn them for audits." It turns installed tools into active behaviors.

2 Slash Commands — Complex Workflows in One Line

Custom commands are markdown files in .claude/commands/. The filename becomes the slash command. The content defines the workflow Claude executes.

Create your own by adding .md files to ~/.claude/commands/. Any workflow you repeat more than twice should be a command.

Permissions — Trust, Not Prompting

The installer configures permissions so you stop clicking "approve" on every operation:

// What gets allowed (all standard tools)
Bash(*), Read, Write, Edit, Glob, Grep, Agent, WebFetch, WebSearch

// What gets hard-blocked
Binary files: mp4, mov, avi, mkv, mp3, wav, flac, zip, tar.gz, dmg...
Secrets:      ~/.ssh/**, ~/.gnupg/**, ~/.aws/**, ~/.azure/**, ~/.kube/**
Credentials:  ~/.docker/config.json, ~/.npmrc, ~/.git-credentials
Crypto keys:  *.pem, *.key, *.gpg, *.cert, *.crt

Full Bash(*) access with the destructive command hook as the guardrail. Binary files denied because Claude can't parse them anyway. Secrets and credentials hard-blocked so they never accidentally end up in a conversation or commit.

Gemini CLI — A Second Brain

The installer optionally sets up Gemini CLI — Google's command-line AI. Gemini Flash is free (60 requests/minute) and brings a different perspective to code review.

Why use two AIs? Because they have different blind spots. Gemini scores 83% on code editing benchmarks where Claude scores 72%. Claude is better at architecture, writing, and tool use. Together, they catch more than either one alone.

The /gemini-review command automates this: Claude writes the code, pipes the diff to Gemini, evaluates the feedback, applies what's valid, and builds to confirm. No manual orchestration.

HOW IT ALL FITS TOGETHER

This isn't a collection of tips. It's a system. Each piece connects:

• Rules define how Claude works — before you type anything
• Hooks enforce boundaries automatically — no approval needed
• Commands encode complex workflows — one slash to execute
• Permissions remove friction for safe operations — block dangerous ones
• Capabilities rule ties it together — Claude knows what it has and uses it proactively

The result: I open Claude Code, give it a task in plain English, and walk away. It knows my standards. It won't destroy anything. It compresses context before it degrades. It reviews its own work with a second AI. It commits when the work is verified.

I stopped using Claude Code as a chat assistant. I use it as an autonomous engineering system with guardrails.

INSTALL IT

Install the Claude Code safeguards from https://github.com/MeriaApp/claude-code-safeguards — clone to /tmp, run install.sh, then delete the clone.

git clone https://github.com/MeriaApp/claude-code-safeguards.git /tmp/claude-code-safeguards
cd /tmp/claude-code-safeguards && ./install.sh
rm -rf /tmp/claude-code-safeguards

Restart Claude Code after installing. Everything activates on the next session.

If you found this useful, I wrote a companion post on every Claude Code feature I actually use — the shortcuts, workflow commands, and daily habits that make the tool feel native.