The strongest point here is that AGENTS.md works best when it captures repository-specific decisions, not generic engineering advice.

We've seen the same pattern with AI coding tools: vague rules like "write clean code" barely change outputs, but concrete examples, forbidden patterns, and architectural constraints do. A short file with "Preferred/Avoid" examples often outperforms a massive document full of principles.

I also like the idea of the five-minute verification test. Removing the context file and comparing diffs is a practical way to measure whether your instructions are actually influencing agent behavior or just adding tokens.

One thing I'd add: include a small "Why this exists" section for non-obvious constraints (e.g., "don't simplify this retry logic because of production rate limits"). Agents are good at following patterns, but preserving past decisions is what prevents them from repeatedly proposing the same rejected ideas.