Claude Code Safeguards

Defensive configuration for Claude Code that prevents crashes, eliminates permission prompting, blocks destructive commands, adds Gemini CLI as a second brain, and installs reusable slash commands.

Built after hitting the 20MB request limit by dropping screenshots into Claude Code — crashing the session with no recovery. These safeguards make that impossible.

Install

Paste this into Claude Code:

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

That's it. Claude handles the rest.

Or manually:

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

Requires: jq (brew install jq / apt install jq)

What's Included

Hooks (3)

Hook	Event	What it does
process-screenshot.sh	PreToolUse/Read	Intercepts image reads, resizes retina screenshots to 1400px, files into project/screenshots/, blocks context blowup
block-destructive-commands.sh	PreToolUse/Bash	Blocks rm -rf, git push --force, sudo, pipe-to-shell attacks, disk-level ops
save-context-before-compact.sh	PreCompact	Injects CONTEXT_STATE.md into context before compression so project state survives

Commands (2)

Slash commands installed to ~/.claude/commands/ — invoke directly in Claude Code:

Command	What it does
/audit <project>	Full engineering audit with parallel agents — code health, architecture, platform compliance. Outputs structured report with P0-P3 severity.
/gemini-review	Delegates current git diff to Gemini CLI for code review. Evaluates findings, applies valid ones, discards false positives, builds to confirm.

Rules (9)

Rule	What it does
binary-file-handling.md	Screenshot protocol, binary file alternatives, context limits
coding-standards.md	Read before edit, minimal diffs, no placeholders, no hardcoded secrets, no temp patches
context-management.md	When to /clear, when to /compact, session hygiene, when to start fresh
file-hygiene.md	_review/ staging system for suspected trash, screenshot/test artifact paths, Desktop rules
full-app-audit.md	Structured audit framework — code health, engineering quality, platform compliance
gemini-orchestration.md	When/how to use Gemini CLI for code review and second opinions
git-workflow.md	Commit only when asked, stage specific files, never force push
quality-standard.md	Verification gate, risk control, root cause fixes, no band-aids
screenshots.md	Screenshot lifecycle — process, file to _used/, auto-cleanup after 7 days
capabilities.md	Meta-rule: tells Claude what's installed and when to use each capability proactively

Permissions

Allow (no more constant prompting):

• Bash, Read, Write, Edit, Glob, Grep, Agent, WebFetch, WebSearch

Deny (hard-blocked):

• Binary files: mp4, mov, avi, mkv, mp3, wav, flac, aac, zip, tar.gz, rar, 7z, dmg, iso
• Design files: psd, ai, sketch, fig
• Secrets: ~/.ssh/**, ~/.gnupg/**, ~/.aws/**, ~/.azure/**, ~/.kube/**
• Credentials: ~/.docker/config.json, ~/.npmrc, ~/.pypirc, ~/.git-credentials, ~/.gem/credentials
• Crypto keys: *.pem, *.key, *.gpg, *.cert, *.crt

Settings

• effortLevel: "high" — better output quality
• $schema reference — editor autocomplete for settings.json

Gemini CLI Setup (Interactive)

The installer offers to:

1. Install Gemini CLI (npm install -g @google/gemini-cli)
2. Open https://aistudio.google.com/apikey in your browser
3. Save your API key to your shell config

Gemini gives Claude a second brain — free code review, 1M context for large codebases, and a different perspective on architecture decisions.

How the Screenshot Hook Works

You drop screenshot
  → Hook intercepts Read
  → Resizes to 1400px wide (retina = 2x)
  → Files to project/screenshots/20260319_170125.png
  → Blocks original Read (exit 2)
  → Tells Claude the filed path via stderr
  → Claude spawns Agent to parse the filed screenshot

Hook exit codes (Claude Code):

Exit	Behavior
0	Allow — tool call proceeds
2	Block — stderr shown to Claude as feedback
1, 3+	Non-blocking error — logged only, tool call still proceeds

How the Destructive Command Hook Works

Blocks these patterns with clear feedback:

Pattern	Block reason
rm -rf / rm -fr	Suggests trash instead
git push --force	Prevents overwriting remote history
git push origin main	Requires feature branch
sudo	Asks user to run manually
curl ... | bash	Download first, review, then execute
mkfs, dd, fdisk	Disk-level ops need manual confirmation

How the Commands Work

Commands are markdown files in ~/.claude/commands/. The filename becomes the slash command name. Contents define the prompt Claude executes.

/audit <project> spawns parallel agents to scan code health, engineering quality, and platform compliance. Outputs a structured report with severity levels (P0-P3) and file:line citations.

/gemini-review captures your git diff, pipes it to Gemini CLI for an independent code review, then evaluates each finding (~30% are false positives), applies the valid ones, and builds to confirm no regressions.

Create your own commands by adding .md files to ~/.claude/commands/.

Customization

Change screenshot max width

Edit hooks/process-screenshot.sh, find 1400:

if [ "$current_width" -gt 1400 ]; then

Change screenshot retention

Default keeps 50. Edit the cleanup line:

ls -t "$screenshots_dir"/* | tail -n +51 | xargs rm -f

Add more deny rules

Edit ~/.claude/settings.json:

"deny": ["Read(*.your-extension)"]

Add your own commands

Create a markdown file in ~/.claude/commands/:

echo 'Do something useful with $ARGUMENTS' > ~/.claude/commands/my-command.md

Then invoke with /my-command whatever args.

Disable a hook

Remove its entry from ~/.claude/settings.json under hooks.PreToolUse or hooks.PreCompact.

Uninstall

cp ~/.claude/settings.json.backup.* ~/.claude/settings.json
rm ~/.claude/hooks/process-screenshot.sh
rm ~/.claude/hooks/block-destructive-commands.sh
rm ~/.claude/hooks/save-context-before-compact.sh
rm ~/.claude/rules/binary-file-handling.md
rm ~/.claude/rules/coding-standards.md
rm ~/.claude/rules/context-management.md
rm ~/.claude/rules/file-hygiene.md
rm ~/.claude/rules/full-app-audit.md
rm ~/.claude/rules/gemini-orchestration.md
rm ~/.claude/rules/git-workflow.md
rm ~/.claude/rules/quality-standard.md
rm ~/.claude/rules/screenshots.md
rm ~/.claude/rules/capabilities.md
rm ~/.claude/commands/audit.md
rm ~/.claude/commands/gemini-review.md

Credits

Built by Jesse Meria and Claude. Inspired by Trail of Bits, Disler's hooks mastery, and the awesome-claude-code community.

License

MIT