You scroll past a label like tech-leads-club / agent-skills [C001] and wonder if it matters to you. If you're only used to chat-style AI tools and you're starting to follow coding agents, this is the part that actually changes your next decision: a giant root AGENTS.md is the most expensive lazy move. Stuffing everything into the root is a tax on every call. [C002]
AGENTS.md is the rules file an agent reads before it works. So when a tiny website task also carries server, release, and security rules from the top of the repo, that is not rigor. It is irrelevant context paid again and again. A tool update is not worth your time because it lists more features. It matters if it changes your next decision.
The evidence here is directional, not exhaustive. One official guide says these files are loaded from the project root down to the folder you are working in. One community example says subfolders can stack their own files. Different places, same lesson: rules closer to the work are usually more valuable than one master file trying to cover everything.
Boundary: this only applies if your project already uses one root AGENTS.md and more AGENTS.md files inside subfolders. There is no token, latency, or model benchmark in the evidence, so the point is not 'this is proven faster.' The point is better rule layering and less context mismatch. If that is your setup, split one bloated root file this week and share this with the person still calling it organized.