RepoGuard Brings Architecture Linting to AI-Generated Code

RepoGuard, a new architecture linting tool, validates AI-generated code from Cursor and Claude against structural and architectural rules your project defines, catching violations that syntax linters and type checkers miss entirely.

As AI code generation becomes a standard part of the development loop, a familiar problem has sharpened: LLMs write code that compiles, passes type checks, and looks reasonable in isolation, but quietly breaks the architectural contracts your codebase depends on. Dependency direction rules, layer boundaries, module ownership, forbidden imports across domains. These are the constraints that keep large codebases maintainable, and they are exactly what a model trained on the open internet has no reason to respect. RepoGuard is a direct response to that problem.

Why it matters

Standard linters (ESLint, Ruff, etc.) enforce style and syntax. Type checkers enforce interface contracts. Neither enforces structural rules like "the payments module must not import from the user profile module" or "all database access must go through the repository layer." When a developer writes code manually, they at least have context about these rules. When Cursor or Claude generates a function, it has none. The result is architectural drift that accumulates silently until it becomes expensive to reverse.

This is not a hypothetical problem. Teams using AI-assisted code generation at scale are already reporting that review burden increases because reviewers have to mentally simulate architectural impact, not just read logic.

Architecture debt from AI-generated code is the same as manual debt, except it compounds at the speed of autocomplete.

What changes in practice

• CI enforcement becomes possible. Architecture rules can now fail a build, not just a code review comment.
• Prompt discipline matters less. Even if your system prompt does not perfectly constrain the model, RepoGuard catches violations downstream.
• Onboarding improves. New developers using AI tools get architecture feedback without needing a senior reviewer to catch every violation.
• Rule documentation becomes executable. Architecture decisions that lived in a wiki or ADR can now be expressed as enforced constraints.

How to use it

1. Define your architecture rules in a config file. Start with the highest-leverage constraints: forbidden cross-module imports, required abstraction layers, banned direct infrastructure access from business logic.
2. Add RepoGuard to your CI pipeline as a required check on pull requests. Treat a failure the same as a failing test.
3. Run it locally in your editor as a pre-commit hook so AI-generated code is validated before it ever leaves the machine.
4. Iterate on rules incrementally. Do not try to encode every architectural decision on day one. Start with the violations that have actually caused incidents.
5. Pair it with prompt engineering guardrails. RepoGuard is a safety net, not a substitute for a well-structured system prompt that tells the model about your module boundaries.

The tool is purpose-built for the Cursor and Claude workflow, which is now the dominant AI-assisted development pattern for teams shipping production software.

Architecture linting is not a new idea, tools like ArchUnit (Java) and Dependency Cruiser (JS) have existed for years. What is new is the urgency: when a model can generate 200 lines in seconds, the gap between generation and architectural review is wide enough to drive a truck through.