Agent Skills open standard

An open, cross-tool format documented at agentskills.io in which a skill directory bundles an entry instructions file with its supporting resources, so any compliant agent can discover and run the skill.

How it works

The standard fixes a portable layout: a skill is a directory whose entry file carries metadata plus instructions, with its supporting resources bundled alongside. A compliant agent reads that layout to discover and load the skill, so the format is the contract rather than any one tool implementation. Adoption is what closes the loop: a directory written against the spec is portable only to the agents that already read the spec the same way.

Why it matters

A shared layout means a skill is authored once and reused wherever a compliant agent runs, rather than rebuilt per tool. The trade-off is that compliance is uneven: the spec declares what a skill is but does not enforce what any one agent must do with it, so partial adherence can be worse than no adherence for authors. A skill that works under one implementation may degrade or fail against another that names itself compliant. Portability holds as a promise pending evidence, not as a guarantee written into the standard.

In practice

A skill directory authored against the standard runs under one compliant agent without changes; the second agent that claims compliance loads the entry file but ignores a primitive the first agent honored, so the same directory degrades silently rather than fails loudly. Authors learn the portability boundary by hitting it.

Practical considerations

The standard fixes a portable layout: an entry file carrying frontmatter metadata (a name and a description, at minimum) plus instructions, with optional bundled subdirectories for executable scripts, reference documentation, and template assets. Loading proceeds by progressive disclosure: a compliant agent reads only the name and description from every available skill at startup, expands the full entry-file instructions when a task matches the description, and executes them while loading bundled resources on demand. The description is the routing surface and doubles as both human documentation and the model's loading-decision input; a description tuned for one role drifts from the other, and a description that does not match how an author would describe the task degrades discoverability without surfacing an error. Portability is transitive only for primitives the standard declares: a skill that depends on a surface one implementation provides natively does not run against another that lacks the primitive, regardless of either agent's compliance claim. Implementation extensions like slash-name invocation hints, autocomplete metadata, or routing opt-outs live above the spec and travel only to the agents that read them.

Defined by Ready Solutions AI