Treat agent skills as software artifacts: single responsibility, interface/implementation separation, and a decision test for skill vs hook vs subagent

arXiv 2607.25032·medium signalXBlueskyLinkedInCopy link

This paper applies ordinary software-engineering discipline to Anthropic's open Agent Skills spec, using Claude Code as the reference implementation. It argues skills should obey single responsibility, separate interface from implementation, and stay loosely coupled while budgeting against a shared token pool via staged content loading — and that the description field is the selection interface, so it must be written for retrieval rather than for humans. The most useful part is a decision framework for choosing between skills, project memory, slash commands, subagents, external tools, and hooks based on what controls execution and what guarantees each mechanism provides, plus a catalog of common faults and third-party trust concerns. It is a position and methodology paper with no empirical evaluation.