The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches and processes Amazon reviews, which are attacker-controllable external data.
Ingestion points: Raw reviews are fetched from Amazon via the ZooData API using the reviews-raw command in scripts/zoodata.py.
Boundary markers: The skill employs a specific prompt template (review-tag-prompt in SKILL.md) that instructs the LLM to "Respond ONLY with a JSON object" and to "Extract only what is EXPLICITLY mentioned", which helps mitigate accidental instruction following.
Capability inventory: The skill is permitted to make network requests to api.zoodata.ai and perform file operations within the monitor-data and /tmp directories via the bundled Python script.
Sanitization: Input review text is truncated to 500 characters in scripts/zoodata.py but does not undergo extensive sanitization before interpolation into the prompt.
[COMMAND_EXECUTION]: The skill executes a bundled Python script (scripts/zoodata.py) to perform API operations and data processing.
Evidence: SKILL.md explicitly lists zoodata.py as an execution target for all API calls. The script is implemented using only the Python standard library, which reduces external dependency and supply chain risks. Furthermore, the script includes a security check to ensure the Bearer token is only transmitted to trusted ZooData hosts or localhost.

amazon-competitor-intelligence-monitor