The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface due to its core function of reading and interacting with external creator platform websites.
Ingestion points: Data from platform web pages (e.g., status messages, current draft state) is read into the agent's context through platform-specific adapter scripts in scripts/v2/platforms/.
Boundary markers: To mitigate the risk of unintended actions triggered by page content, the skill uses a programmatic 'Final Publish Guard' (__VIDEO_PUBLISHER_FINAL_GUARD__) implemented in scripts/v2/ego/core.mjs to block interaction with final publication controls.
Capability inventory: The skill is capable of filesystem operations, browser control via CDP, and spawning subprocesses via node:child_process.
Sanitization: The skill performs extensive validation of content packages and media files before interaction begins, and uses specific regex-based cleaning for topics and tags to ensure they conform to expected formats.
[DYNAMIC_EXECUTION]: The skill utilizes dynamic script assembly to perform browser automation. It concatenates local JavaScript fragments into a larger payload that is then executed within the ego-browser environment.
Evidence: scripts/v2/run-platform.mjs reads core.mjs and platform-specific logic files from the disk and injects them into the browser context.
Mitigation: This behavior is restricted to code shipped within the skill's directory and is essential for interacting with complex, framework-heavy web applications.
[COMMAND_EXECUTION]: The skill uses local command execution to orchestrate the publishing workflow and maintain isolation between platform runners.
Evidence: scripts/v2/publisher.mjs spawns subprocesses to execute run-platform.mjs with specific environment variables and arguments.

video-publisher