When AI Coding Assistants Meet Reality
There's a lot of buzz about how AI can generate entire websites and apps with just a single prompt, and no-code development is being called the future. I've experienced firsthand how AI can significantly improve workflow, making coding faster and more efficient.
However, when it comes to working on code written by multiple developers with different styles and approaches, I've noticed that AI often struggles.
My AI Experiment at ISKCON ISV
I was interning at ISKCON ISV, working on a project that had evolved over time with contributions from multiple developers. The codebase was a mix of old technologies, different coding styles, and a lack of proper documentation, making it quite tricky to navigate.
For my first task, I had to add a new page with specific functionalities. Seemed simple enough, so I decided to let Cursor Pro handle it.
It completely flunked.
The 15,000-Line Problem
The biggest challenge? A 15,000-line API file buried deep in the repo. AI kept getting stuck, hallucinating responses, and repeating mistakes instead of making meaningful progress.
No matter how much I refined the prompts, it just couldn't understand the full context or correctly modify the code.
what AI saw vs reality
Eventually, I gave up on AI and tackled the problem manually. After spending time digging into the existing codebase, I finally got it working.
The Realization
That's when it hit me — this was a relatively small app, but its complexity had grown over time due to multiple developers and outdated practices. Yet, AI still struggled.
What happens when we deal with massive enterprise applications?
Why AI Struggles with Real Codebases
Some argue that AI will keep improving and eventually handle complex coding tasks. But here's the thing:
Real-world codebases are messy
Developers often prioritize making things work over following clean coding principles. AI expects patterns, reality delivers chaos.
Under pressure, clean code goes out the window
Deadlines create technical debt. AI struggles with inconsistencies that humans create under time constraints.
AI can't fully grasp project-specific logic
It might assist in understanding and generating ideas, but it lacks true comprehension of business context and historical decisions.
Where AI Actually Helps
AI is a fantastic tool for:
- Speeding up development on greenfield projects
- Generating boilerplate code for common patterns
- Automating repetitive tasks like writing tests
- Explaining unfamiliar code in smaller chunks
- Brainstorming solutions to well-defined problems
But it isn't ready to take over large-scale, real-world applications just yet.
Final Thoughts
Instead of replacing developers, AI is more like a coding assistant — helpful but not foolproof.
For now, the best devs will be the ones who know how to use AI effectively without fully relying on it.
AI won't replace developers who understand systems.
It will replace developers who don't learn to use it wisely.