The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses a strong directive in the description to override the agent's behavior during pattern selection.
Evidence: The description in SKILL.md begins with "CRITICAL: Use for Robius app architecture patterns," which is a directive designed to force the selection of these specific patterns over others when triggers are met.
[INDIRECT_PROMPT_INJECTION]: The architectural patterns outlined in the skill create a surface for indirect prompt injection through the processing of external data streams.
Ingestion points: Untrusted data enters the application context via AppRequest enums in the worker_task and the MatchEvent trait handlers as shown in SKILL.md.
Boundary markers: The templates do not include markers or instructions to isolate or delimit untrusted data from the rest of the application logic.
Capability inventory: The described patterns include capabilities for writing to the file system (via persistence::save_app_state) and performing network operations (via fetch_data and send_message).
Sanitization: The skill does not provide any examples or guidance on sanitizing external data before it influences the application state or triggers network/file actions.

robius-app-architecture