The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The 'onboard' feature (e.g., onboard diagram-design to https://yoursite.com) instructs the agent to fetch and process content from a remote URL. This creates a surface where malicious instructions embedded in a website's HTML, metadata, or text could attempt to influence the agent's behavior or manipulate the data written to the local style guide.
Ingestion points: Remote website content fetched during the brand onboarding process.
Boundary markers: None explicitly defined in the skill instructions to delimit untrusted data.
Capability inventory: Network read operations (fetching the site) and file write operations (writing to references/style-guide.md).
Sanitization: The skill mentions automatic contrast checks, but does not specify security sanitization for the extracted text or data.
[EXTERNAL_DOWNLOADS]: The skill references external resources including web fonts from Google's official font service (fonts.googleapis.com) and the JSDelivr CDN (cdn.jsdelivr.net) for the Geist font family. It also fetches content from user-specified URLs during style extraction.
[COMMAND_EXECUTION]: The documentation describes several shell commands for installation and usage, such as git clone, ln -s, npx skills add, and npx playwright screenshot. These commands are intended for user execution to manage the skill or export results.
[DYNAMIC_EXECUTION]: The skill's primary function is generating self-contained HTML and SVG files. While the templates provided are static, the ability to generate executable HTML files carries an inherent risk that the agent could be tricked into including malicious scripts if it processes unsanitized adversarial input.

diagram-design-editorial