Builders Are Developing Personal AI Skill Stacks
A thread asking 'how do you manage skills files?' turned into a detailed look at how serious Claude users are building structured knowledge layers for their agents. People described maintaining a .agents/skills/ directory in their dotfiles, using claude.md files per project, keeping company-managed skills that map repo relationships and architecture conventions, and versioning skills alongside code. This is not casual prompt engineering. It is a new form of configuration management.
The pattern: the people getting the most out of coding agents are treating the agent's context as a first-class engineering artifact. One commenter described running jobs unattended 24/7 after March improvements to models and tooling. Another described completely overhauling their repo's skill setup because claude.md was only read at the start of a session, creating gaps during longer tasks.
The key insight from the thread: skills as 'contextual guidance' for a specific author, project, or codebase are unlikely to be made redundant by better base models. The value is in what the model can't know on its own, which is your specific system.
So what?
If you are running a team that uses Claude or similar agents, the teams getting leverage are the ones treating skills and context files as maintained infrastructure, not one-off prompts. This is worth investing in now, before it becomes table stakes. The analogy to ADRs (Architecture Decision Records) that one commenter drew is useful: think of skills files as institutional memory for your AI collaborators.