Ask HN: How Do You Manage Skills Files for AI Agents?

Summary

A Hacker News discussion asks how people discover, organize, validate, and improve “skills” files used by coding agents. Participants broadly distinguish generic skills from project- or team-specific knowledge: general prompting, documentation, and stronger models may make reusable public skills less valuable, while private conventions, deployment procedures, security requirements, and unusual tooling remain difficult for a model to infer. Common practices include keeping skills in Git or dotfiles repositories, distributing them through symlinks, plugins, package-like marketplaces, or synchronization tools, and using short descriptions for progressive disclosure so detailed instructions enter context only when needed. Several participants treat skills as reusable workflow shortcuts, while others combine Markdown guidance with scripts, hooks, MCP tools, or deterministic checks for operations that should not depend on model judgment. Evaluation is another recurring theme: some test skills against failure cases or integration-style evals, whereas others rely on trial and error because formal evaluation can cost more than writing the skill. The thread also raises maintenance and governance concerns, including versioning, review, documentation drift, cross-harness compatibility, and the risk that incorrect shared instructions spread across projects. A minority argues that many skills are merely renamed README files or configuration and that better agents will eventually replace generic ones.