The Agent Skills Directory

[SAFE]: The skill architecture is designed around a 'planner-first' model where the AI agent only generates action plans. Actual execution is offloaded to a browser MCP (Model Context Protocol) and only occurs after explicit user confirmation through a dedicated interface.
[SAFE]: Security controls are explicitly defined to prevent common automation abuses, including policies against password storage, CAPTCHA bypassing, and the automatic expiration of approved actions after 24 hours.
[SAFE]: The skill implements a robust mitigation strategy for indirect prompt injection (Category 8) by ensuring that content observed on web pages cannot authorize or trigger actions. Authorization is strictly tied to the user's out-of-band approval.
Ingestion points: Content from external websites (KDP, AMS, BookBub, ESPs) read via page inspection.
Boundary markers: A mandatory confirmation gate separates planning from execution.
Capability inventory: Uses the browser_drive permission to manipulate web sessions and file_read for local context, with outcome reporting via a POST API.
Sanitization: All observed data is transformed into a 'ConfirmationRequest' for human review before any state-changing actions are taken.

browser-automation