The Agent Skills Directory

[PROMPT_INJECTION]: The skill specifies a localization policy that requires the agent to match the user's input language. This is a functional requirement for global usability and does not exhibit patterns of safety filter bypass or role-play injection.
[COMMAND_EXECUTION]: The skill uses the ae-cli tool to perform analytical queries. The instructions provide precise JSON templates and mandate the use of verified metadata, ensuring that command invocation is constrained to legitimate analytical purposes within the platform ecosystem.
[DATA_EXFILTRATION]: Data output is supported via integration with lark-doc for documentation in Feishu. This represents a standard feature for exporting analytical reports within a corporate environment and does not involve unauthorized external domains.
[INDIRECT_PROMPT_INJECTION]: The skill processes event metadata and user-defined filters to generate reports. While this establishes a data ingestion surface, the risk is mitigated by the platform's reliance on a formal compilation step and structured JSON keys.
Ingestion points: Event names and property values retrieved via analysis-meta or provided by user filters.
Boundary markers: The skill instructs the agent to use exact property names resolved by the compiler and utilizes structured JSON objects for query definitions.
Capability inventory: Execution of analytical queries (ae-cli), creation of dashboards, and writing to external documentation tools (lark-doc).
Sanitization: The agent is explicitly told to resolve names via compiler resolution before execution, rather than accepting raw input directly into aggregation logic.
[DYNAMIC_EXECUTION]: Metric definitions allow for arithmetic formulas (e.g., clicks / impressions). These are domain-specific calculations processed by the ThinkingEngine analytical backend and do not permit arbitrary system-level code execution.

ad-delivery-analysis