Creating larger projects with LLM (as a coder) (medium.com)

A developer challenged the common belief that large-scale projects overwhelm large language models (LLMs), rendering them ineffective beyond small snippets or single files. Despite working on a complex 10,000+ line codebase with multiple interconnected modules, they successfully maintained and extended the project by applying a disciplined, structured workflow that harnesses the LLM as a supportive coding tool rather than a standalone coder. This approach relies heavily on clear upfront design and architecture documentation, iterative development in small, testable increments, and continuous synchronization between evolving docs and code. Key to scaling with LLMs is limiting their scope through frequent context resets and grounding each step in carefully updated DESIGN and ARCHITECTURE files. The developer emphasized systematic review and correction cycles, logging coding rules and guidelines to constrain and improve LLM outputs over time. Frequent commits, refactoring, and a firm division of labor—human-driven planning and decision-making paired with LLM-accelerated code generation and transformation—ensured the project remained navigable, consistent, and adaptable. This workflow highlights that while LLMs don’t “remember” entire large codebases, providing focused, structured context effectively turns them into productive collaborators on sizable software projects, offering an important blueprint for the AI/ML community aiming to integrate LLMs in real-world coding at scale.