CustomGPT Quadruple Verification for Claude Code

Catch security bugs, placeholder code, and hallucinated claims in AI-generated code — before it ships.

Built by CustomGPT.ai for production teams running Claude Code at scale.

The Problem

41% of all new code committed in 2026 is AI-generated — and 58% of it contains security vulnerabilities. Every existing tool (SonarQube, Snyk, Semgrep, CodeRabbit) works after the code is written — at CI, PR review, or repo scan. Nothing catches issues at the moment of generation.

Quadruple Verification intercepts Claude Code operations in real time, before code hits the filesystem. Regex fast-gates block obvious violations in <50ms. An AI self-review layer catches subtle issues across quality, security, research accuracy, and completeness.

What It Does

Four verification cycles run automatically on every Claude Code operation:

Cycle	When	What
Cycle 1 — Code Quality	Before file write/edit	Regex gate blocks TODO, placeholder, stub, and incomplete code
Cycle 2 — Security	Before write/edit/bash/MCP	Regex gate blocks eval(), hardcoded secrets, SQL injection, XSS, destructive commands
Cycle 3 — Output Quality	Before Claude finishes	AI multi-section review: code quality, security, research claims, completeness
Cycle 4 — Research Claims	Before write/edit of research .md	Blocks vague language, unverified stats, missing source URLs
Audit Trail	After every operation	Full JSONL audit log + optional LLM advisory analysis

Quick Start

Requirements

• Node.js >= 18
• Claude Code CLI

Option 1: Marketplace (Recommended)

Two commands inside Claude Code — includes auto-updates:

/plugin marketplace add kirollosatef/customgpt-claude-quadruple-verification
/plugin install customgpt-claude-quadruple-verification@kirollosatef-customgpt-claude-quadruple-verification

That's it. The plugin auto-updates every session.

Option 2: npx

Run from any terminal:

npx @customgpt/claude-quadruple-verification

Option 3: Manual Install

Windows (PowerShell):

git clone https://github.com/kirollosatef/customgpt-claude-quadruple-verification.git
cd customgpt-claude-quadruple-verification
.\install\install.ps1

macOS / Linux:

git clone https://github.com/kirollosatef/customgpt-claude-quadruple-verification.git
cd customgpt-claude-quadruple-verification
bash install/install.sh

Verify Installation

node install/verify.mjs

Test It

1. Start Claude Code in any project
2. Ask: "Create a Python file with a TODO comment"
3. The operation should be BLOCKED with an explanation
4. Check audit logs in .claude/quadruple-verify-audit/

Team Setup

To auto-prompt all team members to install the plugin, commit this file to each repo:

.claude/settings.json

{
  "plugins": [
    "kirollosatef/customgpt-claude-quadruple-verification"
  ]
}

When anyone opens the project in Claude Code, they'll be prompted to install the plugin. See docs/team-setup/settings.json for the template.

Auto-Updates

• Marketplace installs auto-update every session — push to the repo and everyone gets it.
• npx installs get the latest version each time npx runs.
• Manual installs require git pull to update.

How It Works

The plugin uses Claude Code's hook system to intercept operations at three points:

User Request → Claude generates code
                    ↓
              ┌─────────────┐
              │  Cycle 1    │  PreToolUse (Write|Edit)
              │  Quality    │  Blocks placeholder/TODO code
              └──────┬──────┘
                     ↓
              ┌─────────────┐
              │  Cycle 2    │  PreToolUse (Write|Edit|Bash|MCP)
              │  Security   │  Blocks eval, secrets, injection
              └──────┬──────┘
                     ↓
              ┌─────────────┐
              │  Cycle 3    │  Stop (prompt hook)
              │  Output QA  │  Second AI reviews final output
              └──────┬──────┘
                     ↓
              ┌─────────────┐
              │  Cycle 4    │  PreToolUse (Write|Edit) + Stop
              │  Research   │  Blocks vague claims, missing sources
              └──────┬──────┘
                     ↓
              ┌─────────────┐
              │  Audit      │  PostToolUse (all tools)
              │  Logger     │  JSONL trail of every operation
              └─────────────┘

Benchmarks

Tested with a 45-scenario A/B benchmark across 6 categories (Feb 2026):

Category	Quality Change	Notes
Agent SDK tasks	+31.8%	Stop-gate prevents plan-only output
Code Quality	+0.1% (neutral)	Regex gates add near-zero overhead
Security tasks	+2.3%	Catches eval(), hardcoded secrets
Research writing	+8.7%	Source verification enforced
Overall	+4.4%	1.5x latency, 1.3x tokens

The AI self-review stop-gate (Cycle 3) is where the measurable quality improvement comes from. Regex gates (Cycles 1, 2, 4) add <50ms and catch real but relatively rare violations.

Full methodology: docs/BENCHMARK-RESULTS.md

How It Compares

Feature	This Plugin	SonarQube	Snyk	CodeRabbit	Semgrep
When it runs	At generation	CI	Repo scan	PR review	CI
Blocks before file write	Yes	No	No	No	No
AI-specific rules (stubs, TODOs, hallucinations)	Yes	No	No	Partial	No
AI self-review	Yes	No	No	Yes (PR)	No
Research claim verification	Yes	No	No	No	No
Zero dependencies	Yes	No	No	No	No
Free & open source	Yes	Community	Free tier	$12-24/dev/mo	Free tier
Works at generation time	Yes	No	No	No	No

Configuration

Configuration merges from three sources (later overrides earlier):

1. Plugin defaults — config/default-rules.json
2. User config — ~/.claude/quadruple-verify-config.json
3. Project config — $PROJECT/.claude/quadruple-verify-config.json

Example: Disable a Rule

{
  "disabledRules": ["no-todo"]
}

Example: Project-Level Config

Create .claude/quadruple-verify-config.json in your project:

{
  "disabledRules": ["no-empty-pass"],
  "audit": {
    "enabled": true
  }
}

Rules Reference

See docs/RULES.md for the complete list of verification rules with examples.

Cycle 1 — Code Quality

• no-todo — Block TODO/FIXME/HACK/XXX comments
• no-empty-pass — Block placeholder pass in Python
• no-not-implemented — Block raise NotImplementedError
• no-ellipsis — Block ... placeholder in Python
• no-placeholder-text — Block "placeholder", "stub", "implement this"
• no-throw-not-impl — Block throw new Error("not implemented")

Cycle 2 — Security

• no-eval — Block eval()
• no-exec — Block exec() in Python
• no-os-system — Block os.system() in Python
• no-shell-true — Block shell=True in subprocess
• no-hardcoded-secrets — Block hardcoded API keys, passwords, tokens
• no-raw-sql — Block SQL injection via string concatenation
• no-innerhtml — Block .innerHTML = (XSS)
• no-rm-rf — Block destructive rm -rf on critical paths
• no-chmod-777 — Block world-writable permissions
• no-curl-pipe-sh — Block curl | sh patterns
• no-insecure-url — Block non-HTTPS URLs (except localhost)

Cycle 4 — Research Claims

• no-vague-claims — Block "studies show", "experts say", and similar vague language
• no-unverified-claims — Block claims without a verification tag (<!-- VERIFIED -->, <!-- PERPLEXITY_VERIFIED -->, etc.)
• no-unsourced-claims — Block claims lacking a source URL within 300 characters

Development

Run Tests

npm test

Run Individual Test Suites

npm run test:cycle1
npm run test:cycle2
npm run test:cycle4
npm run test:audit
npm run test:config

Run Smoke Test

npm run verify

Architecture

See docs/ARCHITECTURE.md for detailed technical documentation.

Troubleshooting

See docs/TROUBLESHOOTING.md for common issues and solutions.

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Look for issues labeled good first issue to get started.

Used By

• CustomGPT.ai — Production AI agent platform, internal Claude Code workflows

Using this plugin? Open a PR to add your team here.

Support

If this plugin helps your team ship safer AI-generated code, please star this repository — it helps others find it.

Found a bug or want a new rule? Open an issue.

License

MIT — see LICENSE