The Agent Skills Directory

[COMMAND_EXECUTION]: The skill employs Node.js child_process APIs to run its own automation scripts, including the server lifecycle management in scripts/live.mjs, the event-handling loop in scripts/live-poll.mjs, and the variant acceptance logic in scripts/live-accept.mjs.
[COMMAND_EXECUTION]: Helper scripts like scripts/pin.mjs and scripts/cleanup-deprecated.mjs perform file system operations to manage skill installations and clean up legacy files in harness directories (e.g., .agents, .cursor, .claude).
[EXTERNAL_DOWNLOADS]: Reference files such as reference/brand.md instruct the agent to use specific URL patterns for fetching stock imagery from Unsplash (images.unsplash.com) for design mockups when local assets are unavailable.
[DATA_EXFILTRATION]: The live preview feature collects element HTML and styles from the local browser and transmits them to a local server (127.0.0.1). This context is used to generate design alternatives and is not sent to remote third-party domains. The communication is protected by a randomly generated session token.
[PROMPT_INJECTION]: The skill processes code and text extracted from the user's browser, which represents a surface for indirect prompt injection. The agent is instructed to use this data to plan design directions. The skill implements shell-argument escaping in scripts/live-poll.mjs and path-traversal protections in scripts/live-server.mjs as security measures.

impeccable