Did some research and it's a bit more nuanced than this, though still RAG at its core: each skill has a name and brief description that's included verbatim into every prompt, and a Bash "cat" is triggered as a standard tool call to load the full skill specification from disk.

https://platform.claude.com/docs/en/agents-and-tools/agent-s...

And as implemented in Codex: https://github.com/openai/codex/pull/7412/changes#diff-35647...