Skill

Loaded from the host config store, pinned by hash.

Files in the sandbox are never skills, and no agent path writes config: the self-config guard denies every non-read-only call with an argument word holding a .threads path segment.

Fields

nameNamerequired

The name the model passes to load_skill: lowercase letters, digits and underscores, up to 64 characters, unique among the agent's skills.

descriptionstringrequired

One non-empty line, listed with the name in the system prompt so the model knows when to load the skill.

bodystringrequired

The skill's instructions. load_skill adds them to the conversation as trusted instructions; their SHA-256 is pinned in the agent's config.

Edit on GitHub

On this page