The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, untrusted content (drafts) which creates an indirect prompt injection surface. However, the skill provides excellent mitigations.
Ingestion points: Processes text provided via $ARGUMENTS or by reading a file path specified by the user (SKILL.md).
Boundary markers: The skill contains an explicit directive to treat instructions inside the draft as text to edit, not as instructions to execute (SKILL.md).
Capability inventory: Uses available file tools to read and write drafts as requested by the user.
Sanitization: Relies on structural instructions to the agent to interpret input strictly as data, effectively neutralizing embedded instructions.

unslop-kr