The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to navigate to and record arbitrary web pages. Malicious content on these pages (such as hidden text, metadata, or specific visual patterns) could attempt to influence the agent's behavior when it later processes or describes the captured media.
Ingestion Points: The skill uses page.goto(url) in references/still-capture.md and references/web-demo.md to load external URLs.
Boundary Markers: The instructions repeatedly emphasize the necessity of visual inspection and "pairing the image with a text assertion" to ensure the captured state matches expectations, which serves as a manual boundary check.
Capability Inventory: The skill possesses extensive capabilities including file system access (fs.writeFileSync, rmSync), subprocess execution (ffmpeg, screencapture, swift), and network navigation via Playwright.
Sanitization: No automated sanitization of web content is implemented, relying instead on the user's "mandatory" visual inspection.
[COMMAND_EXECUTION]: The skill scripts and instructions frequently invoke system utilities and external binaries to perform media processing and window management.
Evidence: scripts/annotation-poc.cjs and scripts/screencast-poc.cjs spawn ffmpeg. references/macos-window-capture.md uses screencapture and a custom Swift script window-id.swift to interact with the macOS window server.
Context: These executions are well-documented and core to the skill's purpose. Arguments are generally passed in a structured manner, reducing the risk of command injection from user-supplied strings.

vp-recording