If you mostly use chat-style AI and hate finishing a long post without knowing whether it matters to you, this is the part of tech-leads-club / agent-skills worth stopping for. You see the update, almost scroll past, then pause because you do not want to miss the one detail that changes your next move. A root AGENTS.md that tries to do everything is an expensive shortcut. Stuffing everything into the root is paying tax on every call.
That tax is not just about tokens. It is about context mismatch. When frontend, backend, release, and security rules all ride along on every request, the model has to sort through instructions that do not belong to the task in front of it. From a distance that can look thorough. For a beginner trying to save time, it can also turn a simple task into a noisy one.
The evidence here is thin, so the claim needs a boundary. This is a public-docs review, not a hands-on test. Even with that limit, 2 sources and 5 reviewed ideas kept pointing the same way. The official guide describes AGENTS.md loading from the project root down to the current folder and recommends keeping overrides closer to specialized work. A community example made the same practical case: layer smaller task-specific files instead of forcing one master file to explain everything. The point is not just 'save tokens.' The real win is better rule layering.
A tool update is not worth your attention because it lists more features. It is worth it if it changes your next decision. So if you are looking at tech-leads-club / agent-skills, do not ask, 'How many rules can I add?' Ask, 'Which rules should stop riding along on every task?' Keep the root file light, move specific instructions closer to the work, and share this with the person still treating one giant AGENTS.md as a safety blanket.