🤖 AI-Native QA Automation Platform

Built an AI-powered QA automation platform using Claude Code multi-agent orchestration — 4 specialized skill agents (scenario generation, test strategy, test generation, code review) that read domain knowledge, live source code, and business rules before producing output; generated 53 test scenarios stratified across a proper test pyramid (unit / API / component / E2E) for a full-stack Next.js + Express + MySQL application.

Built by Snehal Patil · Senior SDET / AI-Native QA Automation Engineer

---

Business Impact

• Reduced manual test design effort by using specialized AI agents to generate scenarios, assign the right test layer, and produce executable tests.
• Improved test relevance by grounding generation in business rules, live source code, UI selectors, API contracts, and existing tests.
• Enforced a healthy test pyramid by distributing coverage across unit, API, component, and E2E layers instead of overloading UI tests.
• Increased confidence in critical workflows such as authentication, booking rules, seat allocation, and cross-user security through specification-driven validation.

---

What is this?

Most AI-assisted testing tools generate generic tests in isolation. This project is different.

EventHub is a production-quality ticket booking application used as a live system under test. On top of it sits a Claude Code multi-agent QA framework where 4 specialized AI agents collaborate in sequence — each reading domain knowledge, business rules, and actual source code before producing any output.

The result: structured test scenarios, a rationalized test pyramid, self-validating Playwright tests, and AI-driven code review — all grounded in real application behavior, not guesswork.

---

The 4 agents

.claude/skills/
├── create-scenarios/    AI Functional Tester    — generates 53 scenarios across 6 thinking lenses
├── test-strategy/       AI Test Architect       — assigns pyramid layers with source-traced rationale
├── generate-tests/      AI Test Engineer        — writes, runs, diagnoses, and fixes Playwright tests
└── review-tests/        AI Code Reviewer        — audits tests against standards, selectors, domain rules

Agent pipeline

/create-scenarios 
         │
         ▼
  docs/test-scenarios.md        53 scenarios · 6 categories · traced to business rules
         │
         ▼
/test-strategy 
         │
         ▼
  docs/test-strategy.md         layer assignments · rationale · anti-pattern flags
         │
         ▼
/generate-tests 
         │
         ├── reads domain knowledge + source code
         ├── writes test file
         ├── runs tests in real browser (Playwright MCP)
         ├── diagnoses failures: error message + live DOM + source code
         └── fixes and re-runs until green
         │
         ▼
/review-tests 
         │
         └── cross-references best practices · verifies selectors in source · checks domain assertions

What each agent reads before producing output

Agents don’t guess. Each follows a structured read-chain:

---

Test coverage produced

53 scenarios generated by /create-scenarios, stratified by /test-strategy into a proper test pyramid:

Pyramid shape: Unit (5) → API (22) → Component (14) → E2E (12) ✓

Scenarios follow a structured taxonomy:

Defense-in-depth: critical rules covered at multiple layers

Anti-patterns the /test-strategy agent flagged and avoided

---

The system under test

EventHub is a real, production-quality ticket booking application. It gives the agents meaningful complexity: layered architecture, business rules with edge cases, security boundaries, and client-side state machines.

Tech stack

Business rules baked into the agents’ domain knowledge

• Booking limit: max 9 per user — FIFO pruning on overflow; oldest booking deleted first, with same-event fallback that permanently burns a seat
• Booking reference format: [FIRST_CHAR_OF_EVENT_TITLE]-[6_RANDOM_ALPHANUM] — validated at unit + API + E2E layers for defense-in-depth
• Seat atomicity: seat count decrements on booking, restores on cancellation
• Refund eligibility: client-side only — qty = 1 is eligible, qty > 1 is not; 4-second simulated check with spinner state machine
• Security boundary: cross-user booking access returns 403 Access Denied — enforced at service layer, validated at API and E2E layers
• Event limit: max 6 user-created events — FIFO pruning on overflow
• Static events: seeded events are immutable — cannot be edited or deleted

Application pages

---

Project structure

eventhub/
│
├── .claude/skills/
│   ├── create-scenarios/SKILL.md     ← Scenario generation agent
│   ├── test-strategy/SKILL.md        ← Pyramid assignment agent
│   ├── generate-tests/SKILL.md       ← Test writing + self-healing debug agent
│   ├── review-tests/SKILL.md         ← Code review agent
│   ├── eventhub-domain/              ← Domain knowledge (rules, flows, selectors, API)
│   │   ├── SKILL.md
│   │   ├── business-rules.md
│   │   ├── user-flows.md
│   │   ├── ui-selectors.md
│   │   └── api-reference.md
│   └── playwright-best-practices/    ← Coding standards enforced by all agents
│
├── docs/
│   ├── test-scenarios.md             ← 53 generated scenarios (output of /create-scenarios)
│   └── test-strategy.md              ← Layer assignments + rationale (output of /test-strategy)
│
├── tests/
│   └── booking-management.spec.js    ← Playwright E2E suite (output of /generate-tests)
│
├── frontend/                         ← Next.js 14 app (port 3000)
├── backend/                          ← Express API (port 3001)
│   └── src/
│       ├── services/                 ← Business logic — agents read this to assign unit/API layers
│       ├── routes/
│       ├── controllers/
│       ├── repositories/
│       └── validators/
├── playwright.config.ts
└── CLAUDE.md                         ← Project conventions for Claude Code

---

Running locally

Prerequisites

• Node.js 18+
• MySQL 8+

Setup

# Clone and install
git clone 
cd eventhub
npm run setup

# Create the database
mysql -u root -p -e "CREATE DATABASE eventhub CHARACTER SET utf8mb4 COLLATE utf8mb4_unicode_ci;"

# Configure environment
# backend/.env:
#   DATABASE_URL="mysql://root:your_password@localhost:3306/eventhub"
#   PORT=3001
#   NODE_ENV=development
#   CORS_ORIGIN=http://localhost:3000
#
# frontend/.env.local:
#   NEXT_PUBLIC_API_URL=http://localhost:3001/api

# Push schema and seed data
npm run db:push
npm run seed

# Start both servers
npm run dev

Running the tests

npm run test                                                          # all tests, headless
npm run test:ui                                                       # Playwright UI mode
npx playwright test tests/booking-management.spec.js --reporter=line # single file

---

Using the agents

With Claude Code installed and this repo open:

/create-scenarios booking-flow      # generate structured test scenarios
/test-strategy booking              # assign pyramid layers with rationale
/generate-tests booking-flow        # write, run, debug, and fix Playwright tests
/review-tests tests/booking-management.spec.js   # audit against standards and domain rules

Each agent reads its full knowledge chain before producing a single line of output.

---

API reference

Base URL: http://localhost:3001

---

Scripts