How Bundled Skills Work

Skill invocation flow

1. User types /skill-name or Claude’s model decides to invoke a skill

2. Invocability check:

   • user-invocable: false and invoked by user → skill not available in slash menu
   • disable-model-invocation: true and invoked by model → model cannot invoke; only user can
   • Otherwise → continue

3. Feature gate check: If the skill’s isEnabled() returns false (e.g. Anthropic-staff-only skills) → hidden from UI, not invokable

4. Skill prompt loaded from binary (bundled skills) or disk (custom skills)

5. Reference files check:

   • If the skill has a files map → extract files to a temp cache directory; prefix the prompt with Base directory: <dir> so Claude can Read/Grep them
   • If no reference files → skip extraction

6. Skill prompt injected into Claude’s context

7. Claude executes using the tools available in this session

• Registered at startup: Each bundled skill definition calls registerBundledSkill({ name, description, ... }) once when the CLI initializes. The definition compiles into the binary—not markdown files on disk.
• Prompt builder pattern: Unlike built-in commands that execute fixed logic, bundled skills use a getPromptForCommand() function that returns the skill’s instructions as a dynamic prompt. Claude then orchestrates the work using tools.
• Gating via isEnabled(): The optional isEnabled function controls whether a skill appears in the UI and is available to the model. Common checks: USER_TYPE === 'ant' (Anthropic staff only), feature flags, remote-mode availability, or permission policies.
• Reference files extracted to disk: Some bundled skills include a files map (reference documentation, templates, etc.). On first invocation, these files are extracted to a temporary cache directory, and the skill prompt is prefixed with Base directory for this skill: <dir> so Claude can Read/Grep them on demand, just like disk-based skills.
• Invocation control via frontmatter:
  • user-invocable: true (default) — appears in the / menu and can be invoked with /skill-name
  • user-invocable: false — hidden from the / menu; only Claude can invoke via the Skill tool
  • disable-model-invocation: true — Claude cannot invoke; only users can invoke via /skill-name
  • disableModelInvocation: false (default) — Claude can load and invoke the skill automatically when relevant
• Aliases: Some skills have aliases (alternative names). For example, /loop also responds to /proactive.
• Exposed as slash commands: A subset of bundled skills are also listed in the Commands reference, marked as [Skill], and can be invoked directly. Examples: /batch, /simplify, /loop, /debug, /claude-api.

---

← Back to Skills/README.md