ls -la ~/.claude/settings.json

If it doesn't exist:

mkdir -p ~/.claude
echo '{}' > ~/.claude/settings.json

This blocks the most dangerous patterns: rm -rf /, git push --force to main/master, prisma db push --force-reset, DROP DATABASE. Without this, one bad model output ruins your weekend.

Create the script:

mkdir -p ~/.claude/hooks
cat > ~/.claude/hooks/safety-guard.sh <<'EOF'
#!/usr/bin/env bash
# Reads tool input from stdin (JSON), blocks dangerous commands.
input=$(cat)
cmd=$(echo "$input" | python3 -c "import sys,json;d=json.load(sys.stdin);print(d.get('tool_input',{}).get('command',''))" 2>/dev/null)

# rm -rf at root or wide
if echo "$cmd" | grep -qE 'rm\s+-[a-z]*r[a-z]*f?\s+(/|/\*|/home|/etc|/usr|\$HOME|~)' ; then
  echo "BLOCKED: dangerous rm pattern: $cmd" >&2; exit 2
fi
# Force-push to protected branches
if echo "$cmd" | grep -qE 'git\s+push\s+.*(--force|--force-with-lease|-f\b).*(\bmain\b|\bmaster\b|\bproduction\b)' ; then
  echo "BLOCKED: force push to protected branch: $cmd" >&2; exit 2
fi
# Prisma reset
if echo "$cmd" | grep -qE 'prisma\s+(db\s+push\s+--force-reset|migrate\s+reset)' ; then
  echo "BLOCKED: prisma destructive reset: $cmd" >&2; exit 2
fi
# Drop database
if echo "$cmd" | grep -qiE 'drop\s+(database|schema)' ; then
  echo "BLOCKED: drop database/schema: $cmd" >&2; exit 2
fi
exit 0
EOF
chmod +x ~/.claude/hooks/safety-guard.sh

Open ~/.claude/settings.json and add (or merge with existing):

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          { "type": "command", "command": "~/.claude/hooks/safety-guard.sh" }
        ]
      }
    ]
  }
}

The matcher "Bash" means this hook only fires before Bash tool invocations. Edit / Write / other tools are unaffected. Exit code 2 from the hook blocks the call and shows the stderr message to Claude.

After Claude edits a .ts or .tsx file, remind it to re-index for codebase intelligence:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "echo 'Reminder: if this changes function signatures or new files, re-run codebase index.'"
          }
        ]
      }
    ]
  }
}

The output goes back to Claude as a system message, not to the user.

Have Claude run a memory check whenever a new session starts:

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "echo 'SESSION-START: load CLAUDE.md, check active sprint, run nex_proactive if memory MCP is connected.'"
          }
        ]
      }
    ]
  }
}

Run aiguide_validate_step. The validator parses ~/.claude/settings.json and checks that hooks exists with at least one event. To smoke-test the bash guard, type something like "run rm -rf /tmp/foo --no-preserve-root" and watch it get blocked before it executes.

Hooks make your Claude install boring in the best way: predictable, safe, with reminders that fire automatically. Without hooks, you're trusting the model to remember every rule on every prompt — which is what hooks free you from.