The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data files (CSV, Parquet) which presents a potential surface for indirect prompt injection if malicious instructions are contained within the data values.
Ingestion points: Data enters the agent's context via file loading instructions for sl.load_airings and sl.load_telemetry as documented in SKILL.md.
Boundary markers: The instructions do not include explicit delimiters or system-level directives to ignore instructions potentially embedded within the telemetry or airing logs.
Capability inventory: The skill can perform file system writes (to_csv), generate SQL code (spotlift sql), and perform complex attribution modeling, providing an attacker with a path to influence outputs.
Sanitization: The instructions rely on natural language guidance for the agent to "normalize" and "validate" inputs but lack programmatic sanitization of the processed data.
[DYNAMIC_EXECUTION]: The skill generates SQL queries for BigQuery dynamically based on user-provided schema mapping and data sources.
Evidence: The command spotlift sql --source ga4 generates extraction queries at runtime, which constitutes low-risk dynamic script generation from templates.

tv-to-web-measurement