Authored by Mark Seemann, a software architect and consultant, this blog focuses on software design, object-oriented programming, and functional programming principles. Readers can explore articles and discussions on topics such as dependency injection, domain-driven design, and test-driven development. Additionally, they can learn about architectural patterns, design patterns, and best practices for building robust, maintainable software systems.

ploeh blog

LLMs struggle with large codebases due to context window limitations, similar to how human short-term memory constrains programming. The author suggests that architectural patterns like Fractal Architecture—organizing code into small, nested components at every abstraction level—could help both humans and AI systems manage complexity more effectively. These principles from "Code That Fits in Your Head" may be equally valuable for making code more accessible to LLMs.

Code that fits in a context window

I’d recommend this <a href="https://youtu.be/uC44zFz7JSM?si=AZ7dDeJVVbXYBDwX" target="_blank" rel="noopener nofollow">video</a> by Matt Pocock. It’s not the same but a similar topic. They discuss the concept of deep modules making it easier to understand relations between modules which inherently localizes understanding to a smaller section of the codebase. I personally think we should have as little architecture documentation as possible (doesn’t mean zero) and make everything as easily discoverable as possible. We should be able to start a new agent without it having to collect a lot of context wasting tokens along the way.
I think “code that fits in the context window” is a rather weak requirement though. The code required to do a given task should use but a tiny fraction of the context window. Just because the model has a 1M context window doesn’t mean you can just read everything into the conversation and call it a day (for large codebases, it might work for a small hobby project).

This is obviously the medium-term future: But not just smaller encapsulated portions of code, but also there will be a code-gen revolution. Not just LLM-gen’d code, but codegen tools that deterministically produce repeatable and customizable code patterns, that are quantized via DSL’s and other mechanisms that are representable at more efficient higher levels.