The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies a surface where malicious instructions could be embedded in external data and processed by the agent.
Ingestion points: Untrusted data enters the agent context through the ask_user tool during the planning phase, through persona descriptions in the set_roster step, and via the materialFacts and brief fields in the generate_scene call.
Boundary markers: The instructions lack specific guidance on using delimiters or XML tags to wrap untrusted content, which could allow instructions hidden in 'materialFacts' to override the agent's intended behavior during page generation.
Capability inventory: The skill possesses significant capabilities, including persistent data creation via create_stage and generate_scene, and audio synthesis via generate_tts.
Sanitization: There are no explicit requirements for the agent to sanitize, validate, or filter the content provided in personas or attached materials before incorporating them into the final classroom stage.

stage-design