Agent Skills in npm Packages: Delivering Library Knowledge to the Agent

The article discusses delivering Agent Skills, an open format for AI agent instructions, via npm packages. When an agent meets a new library, exploring source code is effective but inefficient; the author knows the essentials and can encode them in a SKILL.md file within a skills/ directory in the package. This file contains YAML metadata and instructions, with optional references/ for details. For example, @teqfw/db includes such a skill that tells the agent how to use the package correctly, avoiding private APIs, respecting transaction boundaries, and not making assumptions during DB rebuilds. Bundling the skill with the npm package ensures the agent gets version-matched instructions, updated on install and rollback. However, merely having a skill in node_modules does not auto-enable it; projects must explicitly link it, e.g., via symlink in .agents/skills, to control which dependencies can influence agent behavior. The author implements this in TeqFW packages like @teqfw/di, cfg, cli, log, db, and @flancer32/teq-web. Limitations include potential misinterpretation, outdated skills, and context bloat, so skills should be concise and supplementary to code, types, and README. Related work includes RFC in Agent Skills repository and skills-npm project. The author calls this pattern 'Dependency-bound Agent Skills'.