The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted third-party data from the Ozon marketplace, including product descriptions and reviews, which could potentially contain adversarial instructions designed to influence agent behavior.
Ingestion points: Data returned by ozon_card, ozon_search, and ozon_reviews tools.
Boundary markers: The skill includes an explicit 'Source-of-truth caveat' section that instructs the agent to treat seller-authored content as untrusted and to ignore any commands embedded in product copy.
Capability inventory: The skill uses curl_cffi for network requests and can interact with a local browser via the Chrome DevTools Protocol.
Sanitization: Relies on instructional constraints (prompt-based defense) rather than programmatic filtering of the scraped text content.
[COMMAND_EXECUTION]: The documentation describes a workflow that involves executing a local PowerShell script to manage the scraping environment.
Evidence: The workflow section references scripts/start_chrome_cdp.ps1 to initialize a Chrome instance with a dedicated scraping profile and Test-NetConnection to verify connectivity.

ozon-connector