The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes raw Amazon customer reviews, which introduces a surface for potential indirect prompt injection from third-party content.
Ingestion points: Raw reviews are retrieved via the realtime/reviews endpoint (executed by zoodata.py reviews-raw).
Boundary markers: The render_review_map_prompt function in scripts/zoodata.py uses triple-quote delimiters (""") to encapsulate untrusted review text, helping the LLM distinguish data from instructions.
Capability inventory: The script is limited to authorized ZooData API communication and local file operations within a private, restricted temporary directory.
Sanitization: Review text is truncated to the first 500 characters before processing to mitigate malicious payload length.
[COMMAND_EXECUTION]: The main script utilizes subprocess calls exclusively for internal maintenance.
Evidence: The _schedule_stale_result_cleanup function in scripts/zoodata.py spawns a detached worker using the current Python interpreter and the script's own path. This is a restricted, self-referential execution pattern used for garbage collection of temporary result directories and is not exploitable for arbitrary command injection.
[SAFE]: The skill provides robust protection for the ZooData API key.
Evidence: The api_call function in scripts/zoodata.py includes a validation check that prevents the Bearer token from being transmitted to any host other than the trusted zoodata.ai domain or localhost.

amazon-listing-audit-pro