The Agent Skills Directory

[PROMPT_INJECTION]: The skill converts arbitrary web content into markdown for agent consumption, which creates a surface for indirect prompt injection. Malicious instructions embedded in a scraped web page could attempt to manipulate the agent's behavior.
Ingestion points: Content is retrieved from the /v1/read endpoint of the Pagewatch service in pagewatch_client.py.
Boundary markers: None implemented; processed markdown is output directly to the agent's context.
Capability inventory: Includes file system writes (tokens, screenshots, PDFs) and outbound network requests to the Pagewatch API.
Sanitization: Relies on the remote service's extraction logic but lacks client-side filtering of potential instructions.
[DATA_EXFILTRATION]: The auxiliary cmd_pdf function in pagewatch_client.py allows reading local files via the --html-file parameter and sending their content to the external API, which could be misused to exfiltrate sensitive information.
[EXTERNAL_DOWNLOADS]: The skill communicates with an external service at https://pagelens.dev for scraping, screenshotting, and PDF generation. This involves transmitting user-supplied URLs and potential file content to a third-party infrastructure.

web-read