You spend hours curating the perfect .claude/skills/ directory. You write meticulous Architecture Decision Records (ADRs) so the AI understands your system. You feel like a master craftsman, guiding a powerful but ignorant apprentice. But there’s a quiet, terrifying realization hiding in the back of your mind: you are building the exact manual that will eventually make you redundant.
If you are building AI-assisted workflows today, you are trapped in an impossible trade-off. Over-invest in organizing your skill files, and you create dead weight the moment the next model update drops. Under-invest, and your AI is clueless, hallucinating, and useless. It is a fragile bridge between human tacit knowledge and machine execution, and the bridge is designed to collapse.
Recently on Hacker News, a developer asked how people manage their AI skills files. The comments revealed a desperate scramble for control. Developers are using chezmoi to manage skills as dotfiles, symlinking .agents/skills/ directories, and writing AGENTS.md notes instructing the AI to log its own “frustrations” during tasks. One user proudly described treating their skills directory as ADRs to help the AI understand the system architecture.
This is brilliant content management. But it is fatal change management.
The more effectively you package your skills for AI, the more you reveal how temporary that packaging is. Your organizational effort is essential, yet obsolete by design.
Most people treat skill files as a static encyclopedia of their expertise. They are not. They are a temporary patch over the model’s current ignorance. As one HN commenter aptly noted, “I believe skills will eventually be eaten by model capabilities.” They are right. The model that needs a custom skill to parse your SSH config today will figure it out autonomously tomorrow. Your perfectly organized directory of community skills—from Mattpocock to Trails of Bits—will vanish into the latent space of a future GPT iteration.
So why do we keep doing it? Because organizing makes us feel safe. It gives us the illusion of control in a rapidly accelerating landscape. We tell ourselves that if we can just structure our knowledge perfectly, we will remain indispensable. But we are confusing the map with the territory.
Curating AI skills is not about preserving your knowledge. It is a desperate bid to delay your own obsolescence.
The scarce resource in the AI era is not the skill itself. It is the ongoing human judgment about what is still worth encoding as models evolve. Any junior developer can write a markdown file documenting a coding convention. It takes a senior engineer to look at a new model release and say, “This entire directory of skills is now dead weight. Delete it.”
If you treat your skill files as permanent assets, you are building a museum to a version of AI that will be dead in six months. If you treat them as disposable scaffolding, you are actually building a competitive advantage. The advantage isn’t in the files; it’s in the human curation layer—the judgment of knowing what the machine still doesn’t know, and ruthlessly pruning what it no longer needs to be told.
The real skill isn’t writing the perfect AI manual. The real skill is knowing exactly when to burn it.
FAQ
Q: If models will just get smarter, why bother with skill files at all?
A: Because right now, today, models are still ignorant of your specific context. Skill files are scaffolding. You don't skip building scaffolding just because the building will eventually stand on its own; you build it so you can work today, and you tear it down the second it's no longer needed.
Q: What should I actually do with my existing skills directory?
A: Stop treating it like a permanent asset. Treat it like a cache. Delete anything that a newer, smarter model could figure out on its own. Keep only the deeply proprietary, tacit knowledge that the model cannot infer from your codebase.
Q: Isn't this just prompt engineering 2.0, destined to die out?
A: Exactly. Skill files are a transitional technology for a transitional era. The people who win won't be the ones with the best skill files; they'll be the ones who recognize the temporary nature of this layer and adapt their workflows faster than the models evolve.