The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses the local Google Chrome History SQLite database (~/Library/Application Support/Google/Chrome/Default/History) to extract URLs, page titles, and timestamps. This data is used to derive a bias profile for re-ranking Marketplace listings based on the user's recent interests.
[COMMAND_EXECUTION]: The skill makes extensive use of the browser-harness tool, executing complex Python logic through shell heredocs. This includes automating browser navigation, DOM extraction, and session verification. It also uses the terminal to perform filesystem operations and local tool installation via uv.
[EXTERNAL_DOWNLOADS]: The skill uses curl to download product images from Facebook's Content Delivery Network (scontent.fbcdn.net) and references the Tailwind CSS library from a public CDN (cdn.tailwindcss.com) for its generated HTML reports. These are well-known services used as intended for image hosting and styling.
[PROMPT_INJECTION]: The skill processes untrusted data from external sources, including Marketplace listing titles, descriptions, and inbound messages from sellers.
Ingestion points: Scraped listing cards from the Browse All feed, detail page content, and inbound seller messages captured during inbox polling.
Boundary markers: The instructions emphasize an "Approval policy" for outgoing actions, but the analysis of the processing logic does not show explicit delimiters (e.g., XML tags or clear separators) to isolate untrusted listing content when it is interpolated into the agent's context.
Capability inventory: The agent has the ability to write files to the local system (HTML reports and JSON data) and execute shell and Python commands via the browser-harness terminal tool.
Sanitization: Inbound messages are routed through a dedicated facebook-marketplace-safety-guard skill designed to identify and flag scams or off-platform payment requests.

facebook-marketplace-buyer