The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides shell command recipes for local execution, using curl and node -e to upload binary files to the WordPress site. These commands are intended to handle media assets that cannot be easily transmitted via the MCP protocol.\n- [EXTERNAL_DOWNLOADS]: The skill uses npx to run external packages like @automattic/mcp-wordpress-remote and tailwindcss from well-known sources. It also includes a tool, wpcanai-sideload-url, which enables the agent to download assets from remote URLs provided by the user and import them into the site's media library.\n- [DYNAMIC_EXECUTION]: Local automation for media uploads is provided as a dynamically executed Node.js script. This script reads from the local file system and performs network requests to the user's configured WordPress endpoint.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external URLs during sideloading and reads site content through tools like wpcanai-read-meta. This ingestion creates an attack surface for indirect prompt injection where malicious instructions embedded in remote files or existing site metadata could influence the agent.\n
Ingestion points: Remote asset URLs (via wpcanai-sideload-url), WordPress site content (via wpcanai-read-meta, wpcanai-list-pages, etc.).\n
Boundary markers: The skill uses Twig comments {# Type / Label #} for structural navigation but does not employ specific security delimiters to isolate untrusted content from the system prompt.\n
Capability inventory: Subprocess execution (npx, curl, node), file system access (node -e using fs.readFileSync), and network operations (fetch, curl, MCP site writes).\n
Sanitization: Content is passed through standard WordPress filters (like wp_kses_post) and Twig syntax parsing guards, which provide some protection but do not fully address adversarial prompt injection.

canai-mcp