Why Your AI Coding Assistant Needs House Rules

I'll admit it: AI coding assistants have changed how I work. Watching one sprint through repetitive code while I focus on the interesting problems is genuinely satisfying. But there's a shadow side that nobody talks about enough. These tools don't have a stop button. Give them access to your repo, and they'll cheerfully rewrite your entire authentication system because you asked them to "clean things up a bit."

That's not a hypothetical. It happens. And it's exactly the problem a tool called anma is trying to solve.

The Problem with "Just Add a README"

Most teams handle this the obvious way: write instructions. CLAUDE.md, CONTRIBUTING.md, some internal wiki page—somewhere there's documentation telling AI agents what to touch and what to leave alone.

The problem? Those documents are invisible when it matters most. When an AI's context window fills up, the first thing to go is the fine print. Your carefully crafted guidelines become suggestions nobody reads.

The result is predictable. Someone asks their AI assistant to add a feature to the auth module. Two hours later, they've got a complete authentication rewrite that completely bypasses the existing session handling. The AI wasn't trying to cause problems—it simply didn't know where the boundaries were.

YAML to the Rescue

anma takes a straightforward approach: define your boundaries in plain YAML files that actually get enforced.

boundaries:
  - scope: auth/
    allow: [read, modify-session]
    deny: [delete, bypass-auth]
  - scope: migrations/
    allow: [read]
    deny: [write, delete]

If you've touched Docker Compose or GitHub Actions, this syntax will feel familiar. The point isn't to be clever—it's to be clear. Anyone on your team can read this and understand exactly what the rules are.

Making Rules Actually Stick

Here's where most documentation approaches fail. They tell, but they don't enforce. anma goes further by compiling your contracts into actual enforcement mechanisms.

Generated CLAUDE.md files: Your contract becomes the official documentation for every AI that touches your project. No hunting for that outdated instruction file—it regenerates automatically.

Pre-commit validation: Before AI-assisted changes land in your repo, anma checks them against the rules. Try to cross a line, and the hook catches it.

CI/CD pipeline checks: Every pull request gets validated, whether the changes came from an AI, a new hire, or a late-night coding session.

This Changes the Team Dynamic

For solo developers, anma is nice to have. For teams, it's practically essential.

Think about onboarding a new AI agent into your workflow. Instead of writing documentation that may or may not get read, you define clear, enforceable boundaries. Your experienced developers set the rules once. Everyone—human and AI—follows them.

This matters especially for:

• Fast-moving startups: Your AI won't undo yesterday's decisions while you're trying to ship today's features.
• Consulting teams: Define client-specific constraints once, then trust that every interaction respects them.
• Larger organizations: Architectural consistency and compliance become automated, not aspirational.

The Real Point

AI coding assistants are going to be everywhere. The real question isn't whether to use them—it's how to use them without shooting yourself in the foot.

Tools like anma represent a sensible middle ground. Not paranoid avoidance, not blind trust—intelligent assistance within actual boundaries.

There's something else worth noting. The YAML approach makes boundary definition accessible. You don't need a custom language or specialized knowledge. The constraints that work best are the ones your entire team can actually read and understand.

Ready to Try It?

If you're already running AI coding assistants—even just for personal projects—setting aside an afternoon to configure anma is worth the effort. You'll sleep better knowing your AI won't surprise you with an unsupervised refactoring session at 2 AM.

The project lives at anma-labs/anma on GitHub. Documentation is clear, and the learning curve is gentle. Think of it as installing guardrails for your increasingly capable AI collaborators.

The future of coding isn't human-only or AI-only. It's both, working together within sensible limits.