New paper applies software engineering to Anthropic Agent Skills

A new paper from Giuseppe Destefanis (arXiv:2607.25032) argues that agent skills—reusable procedural knowledge for LLM agents—should be built like software artifacts. The paper, referencing Anthropic's open Agent Skills specification and Claude Code as the reference implementation, proposes applying classic software-engineering principles such as single responsibility, separation of interface from implementation, low coupling, and economy within a shared token budget. It also calls for behavioral evaluation over deterministic testing. The author details how skills are structured, how their contents load in stages, and how to write the description that determines which skill an agent selects.

The paper places skills in context with other developer mechanisms—project memory files, slash commands, subagents, external tool connections, and hooks—and provides a decision rule based on who decides execution and what guarantees it offers. It outlines an evaluation-driven authoring process, common patterns and faults, and the trust implications of using third-party skills. The work includes UML-style comparisons, a loading model, and an anatomy diagram for skills. This structured approach could help developers build more reliable, reusable AI workflows.