AI Can Write Code, But Can It Maintain It?

Published on February 14, 2025

AI has changed the way we write code. Tools like Claude, GPT-4, and Gemini can generate entire functions, refactor messy code, and even build small applications. But what happens when the project scales beyond a few files?

A recent post from a developer who relied entirely on AI-generated code sheds light on a key issue: AI is great at writing code, but it struggles with long-term project maintenance.

The Experiment: A 30-File AI-Generated Project

The user in question built a Python project with over 30 files using Claude and Cursor. Initially, everything seemed to work fine. But as the project grew:

Claude started forgetting imports
Duplicate logic appeared across multiple files
Bug fixes led to broken functionality
It struggled to optimize code properly

Why Does AI Struggle With Maintenance?

Large-scale software development isn’t just about writing code—it’s about managing complexity.

Context Limitations – Most AI models have a fixed context window. They don’t persist memory between interactions, meaning they lose track of how different files interact.
No Project-Level Awareness – AI operates within isolated chunks of code. Unlike a developer who understands the entire architecture, AI lacks a holistic view of the system.
Surface-Level Fixes – AI-generated bug fixes often don’t address root causes. Instead of refactoring properly, the model deletes lines, introduces new errors, or produces redundant logic.

AI Struggles Even With "Small" Projects

Let’s be clear—30 files is not a large codebase. Many real-world applications have hundreds or thousands of files. If AI is already struggling at this scale, full AI-driven development is still a long way off.

What Can You Do?

Instead of relying on AI to “fix” an entire project, consider a guided approach:

✅ Break down tasks
Instead of asking AI to fix "the entire project," request specific changes in individual files.

✅ Improve context
Provide relevant source files when asking AI to generate or modify code.

✅ Refactor manually
AI can assist, but human review is critical. Ensure that refactored code aligns with best practices.

AI is a Tool, Not a Replacement

AI accelerates development, but it’s not a substitute for software engineering skills. If you don’t understand what AI-generated code does, you’re not writing code—you’re just hoping it works.