The 7 phases of AI-driven development

Summary

The video outlines seven phases of AI-assisted software development, guiding developers from initial idea to final QA. It emphasizes structured workflows using AI coding assistants like Claude Code, including optional research for complex dependencies, prototyping to refine design or architecture, and drafting a PRD to define the end state. Implementation is organized via a canban board, allowing parallel or sequential execution of tasks, followed by a QA loop where humans verify work and generate new tickets for iteration. The approach balances automation and human oversight, focusing on building robust, high-quality, and maintainable applications in the AI era.

Takeaways

• 💡 Start every AI-assisted development project with a clear idea, whether it's an app, feature, bug fix, or code refactor.
• 🔍 Conduct a research phase when dealing with complex external dependencies or APIs, caching findings in a `research.md` file for AI accessibility.
• 🎨 Use prototyping to explore different approaches, especially for UI, architecture, or testing decisions, iterating with human feedback.
• 📝 Create a PRD (Product Requirements Document) to define the intended end state, user experience, and expected system behavior.
• 📋 Convert the PRD into a canban board or ticket list to plan implementation, including task dependencies and opportunities for parallel execution.
• -
• 🤖 Execute the tasks using AI agents, either sequentially or in parallel, with loops like Ralph loops for autonomous iteration.
• ✅ Include a QA phase where AI generates a QA plan and humans review the completed work, creating additional tickets as needed.
• 🔄 The execution and QA phases are iterative, looping multiple times to refine the product to a high-quality state.
• 🛠️ Tools commonly used include Claude Code for AI assistance, GitHub Issues or Linear for ticket management, and Ralph loops for automation.
• 👨‍💻 Human-in-the-loop involvement is crucial for prototyping, QA, and design taste enforcement throughout the process.
• ⚙️ This structured seven-phase approach emphasizes engineering fundamentals for building reliable, long-lasting AI-assisted applications.
• 📈 The methodology is flexible and can evolve, with potential additions like explicit code review as development practices mature.

Q & A

What is the first phase of AI-assisted development according to the video?

The first phase is the Idea phase, where you define the starting point of the project, which can be a full app, a feature, a bug fix, or a code refactor.

Why is a Research phase sometimes necessary?

The Research phase is used when there are external dependencies or difficult-to-access APIs. It caches important information for the AI agent to reference during execution, ensuring the agent doesn't take wrong steps due to lack of context.

How is Prototyping used in AI-assisted development?

Prototyping allows exploration of multiple design or architecture options. It helps impose human taste or preferences on the solution before committing to the PRD. Prototypes are iterated on with human feedback to determine the best approach.

What is a PRD, and what is its purpose in this workflow?

A PRD, or Product Requirements Document, defines the desired end state of the product or feature, including user-facing behavior and high-level implementation notes. It ensures clarity on the final outcome before task execution begins.

How does the Implementation Planning phase work?

The PRD is broken down into tasks on a Kanban board, identifying dependencies and parallelizable work. This makes it possible for AI agents to execute tasks efficiently, either sequentially or in parallel.

What is the Ralph loop mentioned in the video?

The Ralph loop is an execution methodology where AI agents carry out tasks from the Kanban board iteratively with minimal human supervision, allowing for continuous progress and feedback loops.

How does the QA phase interact with the AI execution loop?

In QA, the AI generates a QA plan and humans review the completed work. Any issues identified create new tasks on the Kanban board, looping back into execution until the product reaches the desired quality.

Can code review be part of this workflow?

Yes, while not explicitly mentioned as a separate phase, code review can be included in the QA phase to ensure the code quality meets standards and best practices.

How does the workflow handle large versus small ideas?

The framework is scalable: small ideas like bug fixes follow the same phases as large app development. The process adapts to the scope, whether it requires extensive research or simple sequential execution.

Why might research assets have a limited lifetime?

Research assets can become outdated or inaccurate over time, potentially leading the AI agent to make mistakes. Therefore, they are typically maintained only for the lifetime of the current sprint or idea.

What tools are suggested for managing PRDs and Kanban boards?

GitHub Issues is recommended for both PRDs and task tracking, though tools like Linear may be preferable if you need to represent blocking relationships between tasks.

Why is human involvement emphasized in this AI workflow?

Humans are essential for providing taste, design judgment, QA, and feedback on prototypes. Human review ensures the AI-generated output aligns with expectations and maintains high quality.