Jorge Gomez

Posted on Nov 28, 2025

🚀 How I Used AI, FRDs, Claude Sonnet, Windsurf, and Antigravity to Orchestrate a Complete Backend Without Chaos

A Reproducible Methodology for Generating Real, Clean, and Scalable Software Using LLM Agents Orchestrated Through Formal FRD Documents

Introduction

Over the past few weeks, I’ve been applying a development approach based on FRDs (Functional Requirements Documents) orchestrated with the support of Artificial Intelligence to build modern, clean, and scalable backends. This method allowed me to generate a complete backend—from CRUD, database layer, JWT authentication, migrations, and unit testing—without chaos, without improvisation, and with clean code from the very first commit.

Although FRDs are a traditional standard widely used in software engineering, what’s truly interesting here is not the FRD concept itself, but how FRDs are combined with LLM agents to create a deterministic build process.

This approach aligns with modern techniques such as LLM Orchestration, Task Graphs, and Agentic Pipelines, which are becoming mainstream in tools like Windsurf, Cursor, Claude Sonnet, and VSCode Agents.

One key insight I found early on was that I initially started with a massive FRD—a single document describing all phases (boilerplate, database, authentication, testing, etc.). Even though the document was detailed, two real problems quickly emerged:

LLMs had to process too much information in a single context window, increasing reasoning load and causing the agent to drift, forget steps, or ask redundant questions.
The process became slow because the model entered loops of “confirmation,” uncertainty, ambiguity, and frequent re-validation.

This led me to investigate a more scalable approach: splitting the project into multiple small FRDs, each responsible for a specific phase, and introducing a Master FRD-00 responsible for orchestrating them in the correct order.

With this structure:

The LLM no longer needs to parse a giant document
Each stage has a short, precise FRD
The Master FRD controls the entire execution pipeline
The agent stops asking “what’s next” and simply executes
Adding new features becomes safer and easier because each one lives in its own FRD without interfering with others

What Is FRD-Based Orchestration?

FRDs define:

What needs to be built
In what order
What criteria must be satisfied
Which steps are mandatory
What behaviors the AI must validate
What acceptance conditions each phase must pass

The key insight is that you’re not only describing features, you’re also defining the construction process itself, ensuring accuracy, repeatability, and zero improvisation.

📘 Connection to Existing Practices

FRD-based orchestration does not try to reinvent methodologies—it integrates formal, well-known practices with the new wave of AI-assisted development.

This method incorporates ideas from:

LLM-driven Development Workflows
Prompt Chaining
Task Graph Execution
Reproducible Pipelines
Specification-Driven Engineering

The contribution here lies in how the process is structured, not in the FRD concept itself.

FRD-00: The Orchestration Controller

This master document defines:

Required phases
Clear dependencies
Agent behavior
Validation rules per stage
Strict conditions to advance or stop
Success criteria for each phase

It acts as the logical pipeline that ensures the project can be executed by AI without deviations.

🔗 Ruta

🚀 How Orchestration Actually Begins

Before starting any phase, I execute this initial instruction so the AI loads all FRDs and understands the execution order:

@FRD-00-master-orchestration.md
@FRD-01-boilerplate-core-products.md
@FRD-02-products-database.md
@FRD-03-auth-security.md
@FRD-04-unit-testing.md 

folder name: api-products

FRD-00-master-orchestration.md
FRD-01-boilerplate-core-products.md
FRD-02-products-persistence-typeorm.md
FRD-03-unit-testing.md

Start the orchestration.

Why Is This Critical?

Ensures the agent follows the correct order
Enforces FRD-00 as the master controller
Reduces context overload
Prevents unnecessary loops or clarifying questions
Makes the process reproducible across any AI editor: Windsurf, Antigravity, Cursor, Sonnet, VSCode Agents, etc.

General Project Architecture

The final result was a NestJS backend with:

🧩 Included Modules

Auth (JWT + bcrypt + Passport)
Users (registration, login, hashing)
Products (CRUD with validation)
Centralized config
SQLite + TypeORM + migrations
Unit testing (33 tests)

📁 Final Structure

api-products/
├── src/
│   ├── auth/
│   ├── products/
│   ├── users/
│   ├── config/
│   └── main.ts
├── migrations/
├── README.md
└── package.json

🔥 Phase 1 — Boilerplate + In-Memory CRUD

Project created using nest new
Swagger enabled
Global validation pipes
In-memory CRUD
DTOs with class-validator
Mandatory JSDoc

export class CreateProductDto {
  @IsString()
  name: string;

  @IsBoolean()
  isPremium: boolean;

  @IsNumber()
  price: number;
}

🔗 Ruta

🗄️ Phase 2 — Database + TypeORM

TypeORM integration
Migrations
ProductRepository
Real persistence using SQLite

@Entity()
export class Product {
  @PrimaryGeneratedColumn('uuid')
  id: string;

  @Column()
  name: string;

  @Column()
  isPremium: boolean;

  @Column('float')
  price: number;
}

🔗 Ruta

🔐 Phase 3 — JWT Authentication

UsersService + repository
bcrypt hashing
LocalStrategy + JwtStrategy
JwtAuthGuard
Swagger Bearer Auth

  async login(loginDto: LoginDto): Promise<{ accessToken: string }> {
    const user = await this.usersService.validateCredentials(loginDto.email, loginDto.password);

    if (!user) {
      throw new UnauthorizedException('Invalid credentials');
    }

    const payload = { email: user.email, sub: user.id };
    const accessToken = this.jwtService.sign(payload);

    return { accessToken };
  }

🔗 Ruta

🧪 Phase 4 — Unit Testing

6 test suites
33 tests
Clean mocks
No real database

  describe('login', () => {
    it('should login successfully and return access token', async () => {
      const loginDto: LoginDto = {
        email: 'test@example.com',
        password: 'password123',
      };
      const accessToken = 'jwt-token';
      mockUsersService.validateCredentials.mockResolvedValue(mockUser);
      mockJwtService.sign.mockReturnValue(accessToken);

      const result = await service.login(loginDto);

      expect(result).toEqual({ accessToken });
      expect(usersService.validateCredentials).toHaveBeenCalledWith(loginDto.email, loginDto.password);
      expect(jwtService.sign).toHaveBeenCalledWith({ email: mockUser.email, sub: mockUser.id });
    });

    it('should throw UnauthorizedException for invalid credentials', async () => {
      const loginDto: LoginDto = {
        email: 'test@example.com',
        password: 'wrongpassword',
      };
      mockUsersService.validateCredentials.mockResolvedValue(null);

      await expect(service.login(loginDto)).rejects.toThrow(UnauthorizedException);
      expect(usersService.validateCredentials).toHaveBeenCalledWith(loginDto.email, loginDto.password);
    });
  });

🔗 Ruta

📊 Final Results

All 33 tests passing
Migrations working
JWT fully functional
CRUD operational
Consistent architecture

🖼️ Visual Evidence

(Gallery with screenshots from Windsurf, Antigravity, and Claude Sonnet)

🧠 Why This Matters in the AI Era

AI needs structure
Enables end-to-end project execution
Reinforces best practices in LLM-assisted development workflows.
Reproducible, scalable, modular

🧭 Conclusion

FRD Orchestration is not merely an operational technique—it is a way of structuring collaboration between humans and AI models to produce reproducible, stable, and scalable software.

This method enhances—not replaces—the developer.

📂 Official Repository with FRDs + Full Backend