AIDD Framework

Deep Dive into AI-Driven Development

Lecture 2

Mastering the six-phase workflow for production-ready AI coding

Why We Need AIDD

The gap between AI capability and code quality

Without Methodology

Code works initially, breaks at scale
Duplicate logic scattered everywhere
AI forgets context mid-project
No test coverage
Security vulnerabilities undetected

With AIDD

Production-ready from day one
Single source of truth enforced
Vision document maintains context
TDD guarantees correctness
Review phase catches issues early

Key insight: AI is a powerful tool, but tools need processes. AIDD is that process.

The AIDD Workflow

1

Discover

Map journeys

2

Plan

Create tasks

3

Review

Simplify

4

Execute

TDD build

5

Commit

Push code

6

Test

Validate

Planning Phase

Discover + Plan + Review
Goal: Define exactly what to build

Building Phase

Execute with TDD
Goal: Build with tests first

Validation Phase

Commit + Test
Goal: Verify it works

The Vision Document: Your Project's North Star

Every AIDD project starts with vision.md

Project Overview

What we're building and why it matters

Target Audience

Who uses this and their needs

Primary Goals

Must-achieve objectives

Non-Goals

Explicit scope boundaries

Technical Stack

Languages, frameworks, services

Architecture Decisions

Key choices with reasoning

Constraints

Performance, security, compliance

Success Metrics

How we measure achievement

Why it matters: AI consults this before every task, preventing context drift.

Phase 1: Discover

Discover - Map User Journeys

Read vision.md first. Map user journeys for [feature name]: 1. Personas - Who are the users? What are their roles? 2. Goals - What do they want to achieve? 3. Entry Points - Where do they start? 4. User Flows - What steps do they take? 5. Success Criteria - How do they know it worked? 6. Edge Cases - What could go wrong? Output as a structured story map.

User Personas

Define distinct user types and their characteristics

Story Mapping

Visual representation of user activities

Edge Cases

What happens when things go wrong?

Out of Scope

Explicitly define what we won't build

Phase 2: Plan

Plan - Create Task Epics

Based on the user journeys for [feature], create task epics. For each epic: 1. Title - Clear, actionable name 2. User Story - "As a [user], I want [action] so that [benefit]" 3. Acceptance Criteria - Testable conditions (checkboxes) 4. Dependencies - What must exist first? 5. Complexity - S / M / L estimate Group related tasks. Order by dependency.

Example Epic

Epic: User Registration
Story: As a visitor, I want to create an account so I can access the app
Criteria: [ ] Email validated [ ] Password hashed [ ] Confirmation sent [ ] Error handling
Dependencies: Database schema, Email service | Size: M

Phase 3: Review

Review - Simplify & Validate

Review the task epics for [feature]. Check for: 1. Duplication - Are any requirements repeated across tasks? 2. Coupling - Are tasks too interdependent? 3. Vision Alignment - Does this match vision.md goals & non-goals? 4. Missing Cases - Are error states and edge cases covered? 5. Over-engineering - Can anything be simplified? 6. Security - Are there obvious vulnerabilities? Suggest improvements and flag concerns.

Common Issues

Duplicate validation logic
Missing error states
Scope creep (non-goals)

After Review

Tasks are independent
Single source of truth
Clear test criteria

Phase 4: Execute

Execute - TDD Implementation

Implement [task name] using strict TDD. Step 1: Write a failing test for the first acceptance criterion Step 2: Write minimal code to make the test pass Step 3: Refactor if needed (keep tests green) Step 4: Repeat for next acceptance criterion RULES: - Do NOT implement multiple requirements at once - Show me the test FIRST, wait for approval, then implement - Each commit should be one red-green-refactor cycle

The TDD Cycle

RED: Write failing test → GREEN: Write minimal code to pass → REFACTOR: Clean up while keeping green

Phase 5: Commit

Commit - Push Clean Code

Commit Guidelines

All tests must pass before commit
One logical change per commit
Descriptive commit messages
Reference task/epic in message
No commented-out code

Pre-Commit Checks

Linting (ESLint, Prettier, etc.)
Type checking (TypeScript)
Unit tests (Jest, pytest)
Security scanning (Snyk, etc.)
Code coverage thresholds

Prepare commit for [completed task]: 1. Run all tests and verify they pass 2. Check code against linting rules 3. Write commit message following format: type(scope): description Example: feat(auth): add user registration endpoint

Phase 6: Test

Test - Human + Automated Validation

Human Testing

Think-Aloud Protocol:

User verbalizes thoughts while testing
Reveals usability issues
Catches UX problems AI misses
3-5 users find 65-85% of issues

Automated Testing

E2E Test Scripts:

Playwright/Cypress for UI tests
API integration tests
Performance benchmarks
Screenshot comparisons

Generate test scripts for [feature]: 1. Human Test Script: Step-by-step tasks, questions to ask, what to observe 2. Automated E2E Tests: Happy path + edge cases with assertions

The AIDD Iteration Loop

Continuous refinement, not one-time execution

After Testing

Issues found? Go back to Discover for new requirements or Execute for fixes

Scope Changes

New requirements? Update vision.md, then restart from Discover

Bug Reports

Create new epic, Review for impact, Execute with TDD fix

Refactoring

Review existing code, Execute improvements (tests stay green)

Key principle: The workflow is iterative. Each loop improves quality and understanding.

Choosing the Right AI Tool for Each Phase

Phase	Best Tools	Why
Discover	Claude, ChatGPT	Strong reasoning for requirements analysis
Plan	Claude, Cursor	Task decomposition, codebase awareness
Review	Claude Code, Amazon Q	Code analysis, security scanning
Execute	Claude Code, Cursor	Multi-file edits, TDD support
Commit	GitHub Copilot	Commit message generation
Test	Claude Code, Playwright	Test script generation

Pro tip: Use Claude 4.5 Sonnet or higher for complex reasoning tasks.

Common AIDD Mistakes to Avoid

Skipping Discover

"Just build it" leads to rework. Always map requirements first.

No Vision Document

AI loses context. Vision.md is the source of truth.

Batching Tests

Writing tests after code defeats TDD's purpose.

Skipping Review

Duplicate code and coupling go unnoticed.

No Human Testing

AI can't catch UX issues. Users must validate.

Giant Commits

Hard to review, hard to revert. Small commits win.

AIDD Best Practices

1. Start with Vision

Create vision.md before writing any code. Update it when requirements change.

2. One Task at a Time

Complete one epic fully before starting the next. Avoid context switching.

3. Test First, Always

Write the failing test, show it to AI, then implement. Never batch.

4. Review Everything

AI makes mistakes. Review phase catches them before they become tech debt.

The goal isn't to write code faster. It's to write better code, faster.

AIDD Framework Summary

6

Phases: Discover, Plan, Review, Execute, Commit, Test

1

Vision document as the single source of truth

TDD

Test-Driven Development at the core of Execute

1

Discover

2

Plan

3

Review

4

Execute

5

Commit

6

Test

Questions?

Deep Dive into AIDD Framework

Next: Effective Prompt Engineering

AIDD Framework

Why We Need AIDD

Without Methodology

With AIDD

The AIDD Workflow

Planning Phase

Building Phase

Validation Phase

The Vision Document: Your Project's North Star

Project Overview

Target Audience

Primary Goals

Non-Goals

Technical Stack

Architecture Decisions

Constraints

Success Metrics

Phase 1: Discover

User Personas

Story Mapping

Edge Cases

Out of Scope

Phase 2: Plan

Example Epic

Phase 3: Review

Common Issues

After Review

Phase 4: Execute

The TDD Cycle

Phase 5: Commit

Commit Guidelines

Pre-Commit Checks

Phase 6: Test

Human Testing

Automated Testing

The AIDD Iteration Loop

After Testing

Scope Changes

Bug Reports

Refactoring

Choosing the Right AI Tool for Each Phase

Common AIDD Mistakes to Avoid

Skipping Discover

No Vision Document

Batching Tests

Skipping Review

No Human Testing

Giant Commits

AIDD Best Practices

1. Start with Vision

2. One Task at a Time

3. Test First, Always

4. Review Everything

AIDD Framework Summary

Questions?

Slide Overview