Anthropic Skills

Structured Skill Bundles as a Local Runtime Surface

Anthropic-style skills are valuable because they package reusable operational guidance in a folder that can be inspected, shared, and maintained as a coherent unit. The practical benefit is not merely stylistic consistency. It is that instructions, supporting files, and limited local procedures can be kept together under one stable skill boundary.

In AchillesAgentLib, this family aligns with that external standard but is not treated as a passive documentation bundle. The descriptor under SKILL.md becomes the main behavioral surface for a loop session. The skill body is therefore not only descriptive text. It functions as live session guidance inside the runtime.

This is why Anthropic Skills matter architecturally. They show that even an apparently simple instruction bundle already behaves like a localized interpreter of task intent. The library preserves the external packaging intuition while making the execution path explicit and stateful inside its own session model.

How the Subsystem Works in Practice

An Anthropic Skill is discovered through SKILL.md, not skill.md. During preparation, the subsystem records the parsed descriptor together with the files found under scripts/ and resources/. At execution time, it starts or resumes a loop session whose system prompt is derived from the raw skill body and the current project root. The authoring implications of this descriptor format, including frontmatter support and optional helper folders, are described in Authoring.

The runtime tool surface is then built from a bounded set of internal skills and from two optional filesystem-backed tools: run-script when scripts/ exists, and get-resource when resources/ exists. The subsystem therefore does not expose an arbitrary local execution surface. It constructs a limited tool environment around the skill bundle and lets the loop session use that environment over multiple turns. The exact runtime behavior is described in Runtime.