The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content by navigating to user-provided URLs using a Chromium instance. While this presents an attack surface where a malicious website could attempt to influence the agent via content, the risk is minimal as the skill uses specific, hardcoded JavaScript 'recipes' to extract only calculated CSS properties (e.g., background-color, border-radius) rather than general text processing.
Ingestion points: The navigate_page tool is used to load user-provided URLs in SKILL.md.
Boundary markers: Explicit instructions in SKILL.md tell the agent to focus solely on visual tokens and ignore the actual site content/copy.
Capability inventory: The skill utilizes read_file, write_file, navigate_page, and evaluate_script via the chrome-devtools MCP server.
Sanitization: Extraction is limited to computed styles via the scripts in extraction-recipes.md, which return structured data for color conversion to OKLCH.
[COMMAND_EXECUTION]: The skill performs file system operations by reading and patching the project's CSS files (e.g., globals.css). This is the intended primary purpose of the skill and includes a requirement to show a diff to the user before overwriting custom themes.
[DYNAMIC_EXECUTION]: The skill uses evaluate_script to execute JavaScript within the target browser page. These scripts are statically defined in extraction-recipes.md and are used for data extraction and theme toggling rather than executing untrusted logic generated at runtime.

extract-theme