The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources such as social media bios, video titles, and profile metadata which could contain malicious instructions.
Ingestion points: Data enters the agent context through the YouTube API script and oEmbed URLs for platforms like TikTok and X.
Boundary markers: The instructions do not define explicit delimiters to isolate external content from the prompt templates.
Capability inventory: The skill maintains permissions to write to the local file system (under the memory directory) and execute local scripts.
Sanitization: No specific sanitization or escaping of the ingested social media content is described.
[COMMAND_EXECUTION]: The skill uses shell commands to invoke local Python scripts with parameters provided by the user or sourced externally.
Evidence: The skill executes python3 "${CLAUDE_PLUGIN_ROOT}/scripts/connectors/youtube.py" channel @handle.
Risk: Influencer handles are passed directly as arguments to shell commands; if a handle contains shell metacharacters, it could lead to command injection within the local execution environment.
[EXTERNAL_DOWNLOADS]: The skill fetches metadata from well-known social media services.
Evidence: The skill makes requests to youtube.com, tiktok.com, and twitter.com to retrieve oEmbed and API data.

influencer-discovery