Skills are expected to be atomic. We need to be able to redirect an agent to a different skill to complement.
Skills are expected to be atomic by design, but real workflows require composition. One skill often needs to delegate to or reference another. There's currently no convention for cross-skill references, and no clear boundary between the skill-resolution mechanism and the agent's policy for when/how to follow those references.

Proposal:

• Naming convention for skill references: Standardize on owner/repo@skill:version (already used by find-skills). This gives us namespacing, versioning potential, and discoverability.
• Explicit dependency declaration: Skills should be able to declare complementary skills in their metadata (e.g. a complements or depends field in SKILL.md frontmatter), rather than relying on implicit agent behavior. I would choose a natural language reference (use X when Y).
• Separation of concerns: Skill resolution (finding + loading a referenced skill) is a conventional concern. Whether to automatically follow a cross-skill reference is an agent policy concern. These must stay decoupled: a skill says "you may also need X," the agent decides whether to fetch, ask the human, and apply X.