Analyzed: March 31, 2026 leak snapshot

​
Skill sources

src/skills/loadSkillsDir.ts supports skills from several sources:
  • user settings home directory
  • project .claude/skills
  • managed policy path
  • plugins
  • bundled skills
  • MCP skill builders

​
File format

Skills are markdown-driven. The loader parses frontmatter and markdown content, then derives command-like metadata such as:
  • display name
  • description
  • allowed tools
  • argument names
  • when_to_use
  • model override
  • effort
  • hooks
  • execution context

​
Token budgeting

estimateSkillFrontmatterTokens() estimates skill cost from frontmatter only. That indicates the loader is intentionally trying to keep the main context small until the skill is actually invoked.

​
Paths and scoping

The loader supports path-scoped skills through frontmatter parsing. It also resolves symlinks via realpath when deduplicating files, which helps avoid the same skill being loaded twice through overlapping paths.

​
User-invocable versus model-invocable

Frontmatter can mark a skill as user-invocable or disable model invocation. That is an important distinction: not every loaded skill is intended to be callable by the model in the same way.