If you mostly use chat-style AI and keep wondering whether every new tool deserves your attention, this is the part that matters. What is not written into a rules file does not exist for AI. That is the real takeaway from addyosmani / agent-skills.
A release is worth your time not because it lists more features, but because it changes your next decision. If you only watch the surface-level excitement, you can waste time, budget, and attention in the wrong place. The quieter cost is that your repeatable judgment stays trapped in people's heads, so the AI keeps missing the same preferences and making the same avoidable mistakes.
The public guide makes the case clearly. It puts rules files at context layer #1 and treats them as the highest-leverage form of persistent context, with examples like CLAUDE.md and AGENTS.md [C001]. In plain English, a rules file is just the written version of 'how we do things here' so the AI does not have to guess every time.
The same guide also pushes back on a comforting myth: the agent will not simply absorb your team norms on its own. If it is not written down, it effectively does not exist for the AI [C001]. That is why 'CLAUDE.md beats a model upgrade' is a useful provocation. Better models can still matter, but a newer model cannot use knowledge you never externalized.
Next step: before you tweak prompts again or pay for another upgrade, write down the rules you keep repeating anyway. What good output looks like. What to avoid. What 'done' means. Then share this with the person who still treats AI quality as a pure model-shopping problem. Based on the public GitHub guide, not a live coding test.