The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to read content from files outside its immediate directory, which could contain instructions that influence the agent's behavior without proper sanitization.
Ingestion points: Reads ../satisfy-your-agent/SKILL.md and ../satisfy-your-agent/references/commands.md in SKILL.md.
Boundary markers: Absent. There are no instructions to ignore embedded commands or treat the external content as inert data.
Capability inventory: The agent maintains access to its standard toolset while processing these external files.
Sanitization: Absent. Content is loaded directly into the context.
[COMMAND_EXECUTION]: The instruction to "Pick one very small playful thing for yourself, do it immediately" is highly open-ended. While intended for harmless role-play, it provides a directive for autonomous action that could lead the agent to execute commands or access network resources if not properly constrained by the environment's security posture.

sya-snack