The Standard

The Advanced Skill Library Standard defines how to structure a plugin as an advanced skill library that goes beyond a flat collection of standalone skills.

The two-axis model

• Structure: a component (the atomic building block and unit of reuse) sits inside a plugin (the package with a manifest and the one version, the unit of release), which a marketplace catalogs.
• Quality: a skill library is not a separate artifact. It is the grade a plugin earns by conforming - Bronze, Silver, or Gold.

So the path is: loose components, then a plugin, then a graded library.

Components

The Standard specifies each component type: skills, slash commands, subagents, hooks, workflows, chain contracts, MCP servers, and the AGENTS.md instructions entrypoint. Each has a required structure, a frontmatter contract, per-agent format notes, and validation rules.

Design principles

• Deterministic where objective; LLM where judgment. Conformance, portability, and security are objective, so they are checked by portable scripts with no model in the loop (CI-safe). Domain fit, output quality, and triggering behavior are judgment, so they are delegated to an opt-in LLM layer that sits beside the gate and never returns a CI pass or fail.
• One dispatcher, builders per type. A single evaluate is the assessment front door; a build-<type> skill authors each component type with create and improve modes.
• A la carte minimalism. A plugin ships only what it needs. Governance files (a chain contract, for instance) are required only when the thing they govern is in use.

Cross-agent, honestly

The Standard is concept-convergent and format-divergent across Claude Code and Codex. Skills, MCP, and AGENTS.md are portable on both. A command’s Codex form is its backing skill. Subagents, output styles, and statuslines are Claude-only for plugin distribution. The toolkit encodes these facts so a plugin’s claims match what the agents actually support.