Model invocation

Model invocation is Claude loading a skill on its own because your request matched the skill's description, as opposed to you typing /skill-name. By default, both methods are available: you can type /skill-name to invoke it directly, and Claude can load it automatically when relevant.

Two frontmatter fields in the SKILL.md file restrict this behavior. Setting disable-model-invocation: true means only you can invoke the skill directly; Claude will not load it automatically. In this configuration, the skill's description is not included in Claude's context when you invoke it directly. Conversely, user-invocable: false means only Claude can invoke the skill automatically. The skill's description is always provided to Claude in this case. It's important to note the asymmetry: user-invocable controls menu visibility only, not the ability for Claude to use the skill programmatically. To block programmatic invocation, you must use disable-model-invocation: true.

Automatic invocation can also be restricted by file path. The paths frontmatter field accepts glob patterns. When paths is set, Claude will only load the skill automatically when it is working with files that match these patterns.

Because model invocation relies on Claude's interpretation of the skill's description text, it is the least deterministic aspect of the system. This is why SkillProof evaluates the accuracy of automatic triggering as a distinct criterion, rather than assuming a skill will fire simply because it has been installed. This contrasts with direct invocation, which is governed by the skill-description field and explicit user commands.