⚡️ Matt Pocock - Why Engineering Fundamentals matter MORE now
Matt Pocock of AI Hero discusses the critical role of classic software engineering principles in the new era of AI development. He explores how concepts like Domain-Driven Design (DDD), deep modules, and intentional architecture are essential for building maintainable systems with AI, and shares his unique teaching philosophy for the rapidly evolving field of AI Engineering.
Matt Pocock, known for his work in the TypeScript community, has transitioned his focus to AI engineering with his new venture, AI Hero. He emphasizes that as AI tooling advances, the fundamentals of software engineering are not becoming obsolete but are, in fact, more critical than ever.
The Fallacy of “Code Doesn’t Matter”
A prevalent idea in the AI space is that code is merely a “compile target” for natural language prompts, and developers no longer need to care about its structure. Pocock argues strongly against this, based on his own experiences.
“Every time I tried that, every time I tried to ignore the code, I would just end up with a this terrible mess.”
He found that the key to improving the AI “compiler” was to revisit classic software engineering texts like Extreme Programming, The Pragmatic Programmer, and A Philosophy of Software Design. The core insight is that principles designed to make code manageable for humans are equally effective for AI. A codebase that is easy for a human to change is also easy for an AI to change and reason about. This means being intentional about architecture and modules yields significant dividends in development speed and maintainability.
System Design: Deep Modules and the “Narrow Waist”
The conversation highlights the importance of well-defined architectural patterns for containing complexity when working with AI. This is framed through two key concepts:
- Deep Modules: A concept from John Ousterhout’s A Philosophy of Software Design, a deep module is characterized by a large amount of functionality hidden behind a simple interface. This is analogous to the ports and adapters pattern.
- The “Narrow Waist”: Borrowed from internet architecture, this concept involves creating well-defined interfaces that contain “slop” or complexity. The developer defines what goes in and what comes out, and can then delegate the internal implementation to an AI without losing control of the overall system.
This approach allows developers to be intentional about the high-level design while leveraging AI for the detailed, boilerplate implementation inside these self-contained boxes.
Domain-Driven Design (DDD) for a Shared Language with AI
Domain-Driven Design (DDD) is presented as a powerful framework for aligning human developers and AI models. Because DDD has been around for a long time, its concepts are well-represented in the latent space of large language models.
The central goal of DDD is to align the code with the business language, creating a Ubiquitous Language. This is incredibly valuable for AI-driven development.
“You’re trying to kind of create a unified domain model so that… the AI and you are speaking the same language.”
Pocock has put this into practice with a popular skill in his mattpocock/skills repository called ubiquitous-language-skill. This tool scans a codebase, helps define the core domain terms in a markdown document, and keeps it accessible. By keeping this document open while prompting, a developer can ensure clarity and consistency, leading to much better interactions and outputs from the AI, as it will use the correct domain-specific terminology when exploring the repository.
The Future of Coding Agents and Inversion of Control
Pocock observes that the architecture of coding agents is beginning to converge around similar patterns. He predicts that the next significant development will be around the inversion of control (IoC), giving more power back to the developer. This involves a trade-off:
- Black-Box Harnesses (e.g., early Claude Code): These are easy to use but offer low observability and control. You trust the system and work within its constraints.
- Primitive-Based Frameworks: These provide total control and observability over every part of the system but require the developer to maintain more code and manage more complexity.
Exploring this trade-off between ease of use and granular control will be a key theme in the evolution of AI development tools.
On Teaching and Learning in the AI Era
With a background as a voice coach, Pocock brings a unique perspective to education. His course creation process is a two-month “explore and exploit” cycle, starting with a Zettelkasten of ideas in Obsidian and culminating in a structured curriculum where each lesson teaches only one thing.
He categorizes learning into three types:
- Knowledge: Taught through lectures.
- Skills: Taught through interactive exercises.
- Wisdom: The hardest to teach, traditionally done via small group discussions.
Interestingly, he notes that AI offers a new path to wisdom by allowing a user to “talk” through concepts. However, he has found that his audience still prefers traditional learning formats (lectures, structured exercises) and that leaning too heavily into experimental AI-based teaching methods can turn people off.