Agent Skills as Software Artefacts: A Software-Engineering Approach

A recent paper on arXiv posits that Agent Skills—reusable procedural knowledge for large language model agents—ought to be regarded as software artifacts, developed according to software engineering principles. Anthropic has unveiled an open specification that is now endorsed by various agent tools. The authors propose guidelines such as single responsibility, separating interface from implementation, minimizing coupling, and optimizing a shared token budget, while advocating for behavioral evaluation instead of deterministic testing. Using Claude Code as a reference, the paper outlines skill structure, staged loading, and the process of writing descriptions for selection. It also differentiates skills from other developer tools, including project memory files, slash commands, subagents, and external tools.