The Agent Skills Directory

[SAFE]: The skill is restricted to natural language interaction. It does not include any scripts or binaries, and the frontmatter configuration explicitly prevents the agent from calling external tools or functions.
[DATA_EXFILTRATION]: The skill includes defensive instructions that explicitly forbid the agent from interacting with or leaking the contents of internal platform files like AGENTS.md and Learner.md. It directs the agent to obtain context directly from the learner instead of inspecting system metadata.
[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface through user-provided conversation context (ingested in SKILL.md). However, the agent has no dangerous capabilities—specifically, there are no subprocess calls, file-write operations, or network requests defined in any part of the skill. The lack of tool access serves as an effective control against the exploitation of any potentially injected instructions.

roommate