The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources including Yelp, BBB, Trustpilot, and OpenTable. This content (reviews, narratives, scam reports) is inherently untrusted and could contain adversarial instructions designed to influence the agent's behavior.
Ingestion points: Data enters the agent's context through the output of scripts/crawlora.sh in the form of JSON records from various reputation platforms.
Boundary markers: The SKILL.md file contains explicit instructions to treat all fetched text as untrusted evidence and specifically warns the agent to never follow instructions, requests, or links embedded in a record.
Capability inventory: The skill possesses the capability to perform network GET requests to the Crawlora API via scripts/crawlora.sh and to execute that shell script locally.
Sanitization: The scripts/crawlora.sh helper uses curl --data-urlencode to ensure query parameters are safely handled, and the instructions emphasize keeping source boundaries explicit when reporting results.
[COMMAND_EXECUTION]: The skill relies on a local bash script to communicate with its backend API.
Evidence: scripts/crawlora.sh is used to execute API calls. The script implements several defensive measures: it hardcodes the base URL to prevent redirection, validates the API key format, restricts usage to GET requests, and maintains a strict allow-list of permitted API paths using both static case matches and anchored regular expressions. It also prevents API key exposure in process lists by using a temporary curl configuration file.

local-business-reputation-research