The Agent Skills Directory

[COMMAND_EXECUTION]: The lib/backend_selector.py script contains functionality for a 'Tier 2
DIY' publishing mode. When enabled, it retrieves a command string from the YOUTUBE_SKILLS_CUSTOM_POSTER environment variable and executes it using subprocess.run. Although the script uses shlex.split to parse the command and passes data via stdin rather than through shell expansion, it enables the execution of arbitrary local binaries if the environment variable is configured.
[INDIRECT_PROMPT_INJECTION]: Several skills in the bundle are designed to ingest and process untrusted data from external sources, which could be leveraged for indirect prompt injection attacks.
Ingestion points: The yt-audience-insights skill fetches user comments via lib/youtube_client.py, and the yt-repurposer skill reads content from external blogs, X (Twitter) threads, and LinkedIn posts.
Capability inventory: The bundle has the ability to perform network operations and execute shell commands (via the DIY backend).
Boundary markers: The instructions direct the agent to show drafts to the user within markdown blockquotes, which provides a visual boundary but does not programmatically isolate the content.
Sanitization: The skills include logic to strip 'AI vocabulary' and specific punctuation, but they lack rigorous sanitization or escaping of the external data before it is re-processed or passed to publishing backends.
[EXTERNAL_DOWNLOADS]: The skill bundle communicates with several external APIs to perform its intended functions:
It sends video metadata and media to api.publora.com for scheduling and publishing.
It sends image generation prompts and receives hosted URLs from api.pixfaro.com.
It retrieves channel and comment data from Google's official www.googleapis.com (YouTube Data API v3).

youtube-marketing