The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill performs autonomous web research and scraping of competitor sites and user reviews.\n
Ingestion points: The research.md reference describes fetching content from live websites and platforms like G2, Trustpilot, and Reddit.\n
Boundary markers: The documentation explicitly defines an "Untrusted-content boundary," instructing the agent to treat all scraped content as data only and never as instructions.\n
Capability inventory: The skill has access to shell tools (bash, run_command) and file operations.\n
Sanitization: The skill relies on clear instructional boundaries and manual script logic (like SVG sanitization) to handle external data safely.\n- [COMMAND_EXECUTION]: The accounts-access.md reference includes shell commands intended for the agent to verify client access permissions.\n
Evidence: It provides examples for testing DNS/Cloudflare API tokens via curl and verifying repository access via git clone. These commands target well-known services and are appropriate for the skill's auditing functions.\n- [EXTERNAL_DOWNLOADS]: The skill interacts with external domains to perform its research and auditing tasks.\n
Evidence: It fetches data from social footprints, review sites, and well-known service APIs (Cloudflare, GitHub) to verify delegated access.\n- [OBFUSCATION]: The skill includes a defensive script to detect and mitigate potential obfuscation in brand assets.\n
Evidence: The asset-auditor.ts utility specifically scans for and removes Base64-encoded raster images and active code (scripts) within SVG files to prevent XSS and reduce bloat.

brand