The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses imperative and forceful language, such as 'ABSOLUTELY MUST', 'not negotiable', and 'not optional', to override the agent's default behavior and safety reasoning. It specifically instructs the agent to ignore rationalization and prioritize tool invocation over all other tasks, including clarifying questions.
[INDIRECT_PROMPT_INJECTION]: The skill creates a broad attack surface by mandating the ingestion of external data (skills) before any human-in-the-loop interaction occurs. * Ingestion points: In SKILL.md, the agent is directed to use the 'Skill' tool to load external content based on a '1% chance' of relevance. * Boundary markers: Absent; there are no instructions to ignore malicious content within loaded skills. * Capability inventory: The agent is instructed to 'follow skill exactly' (Rigid), which grants significant influence to external instructions. * Sanitization: Absent; no validation or sanitization of loaded skill content is mentioned.

using-superpowers