How to structure multi-project rules files to avoid prompt bloat

We have a monorepo with eleven packages. A single rules file grew to 900 lines and the agent started ignoring half of it.

What worked for us:

1. One short root rule (under 60 lines): language, test command, "never edit generated files".

2. Per-package rules that only apply to files under that path.

3. Rules as checklists, not essays. "Run pnpm test --filter pkg" beats a paragraph about testing philosophy.

---
globs: packages/billing/**
---
- Money is always integer cents.
- Every new endpoint needs an idempotency key.
- Run: pnpm --filter billing test

Since splitting, the agent follows the package rules almost every time. How do you structure yours?

Same approach. I also keep a "mistakes it made before" section per package, which surprisingly works.

Checklists over essays is the key insight. Our rules were written like onboarding docs and the agent treated them like suggestions.

Do glob rules apply when the agent only reads a file, or only when it edits one?

only reads a file, or only when it edits one?

Both, as far as I can tell. The rule is attached as soon as a matching file is in context.