I've been using AI coding assistants like Copilot and Cursor a lot lately, and while they do help me be more productive, I'm running into a serious issue: they produce working code that ultimately damages my project's architecture. I've noticed problems such as UI components directly importing database or infrastructure code, domain logic relying on infrastructure, and even AI 'refactoring' things like security and payment logic incorrectly. Deprecated dependencies are creeping in too, and often these architectural issues only surface later during code reviews. Unfortunately, linting, TypeScript, and tests aren't catching these problems, as they typically focus on syntax and behavior rather than architectural intent.
To tackle this, I thought, what if AI-generated code had to pass an architectural contract before being accepted? So, I developed a tool called Intent-Guard, which is a CLI application that allows you to define architectural rules, validates AI-generated code against those rules, and fails fast if there's any architectural violation. It doesn't generate or fix code, nor does it replace ESLint or TypeScript; it merely checks for structure and intent. I'm in the early stages of this project and would love to hear honest feedback.
So, I'm curious: Are you facing similar architectural struggles with AI-generated code? Would a guard layer like Intent-Guard be helpful or just unnecessary? What edge cases do you think could make this tool ineffective? And how would you solve this issue? I'm open to sharing more details or examples if anyone's interested!
2 Answers
I think your idea might be overkill. It seems like you’re trying to automate a level of understanding that’s pretty complex. If you keep relying on AI without engaging critically with it, you might have bigger problems than just lacking architectural oversight.
I get where you're coming from—the issues you've outlined are definitely real. I’ve also faced a lot of repetitive linting and TypeScript errors. What worked for me is creating a document with project rules I include in every prompt for AI. Plus, I added some checks in my CI/CD. It’s still a work in progress, though.
That's a nice approach! It really helps to put these rules down in a project rules file. I'm curious, where do you find your method breaking down? Is it in coverage, or in the AI ignoring constraints?