The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to run local Python scripts, specifically tools/gen_thumbnail.py for image generation and tools/yt_stats.py for retrieving video analytics.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input during its intake phase (Stage 1) to build descriptive prompts for the image generation model, which could be exploited to influence the generated output.
Ingestion points: SKILL.md Stage 1 intake captures the video's "Core idea" and "concrete takeaway."
Boundary markers: The prompt templates defined in references/thumbnail-generation.md do not utilize explicit delimiters to separate user input from the structural instructions for the image model.
Capability inventory: The skill uses tools/gen_thumbnail.py to generate image files based on these prompts.
Sanitization: The instructions include a "verify loop" in references/thumbnail-generation.md that directs the agent to inspect the resulting images for text accuracy, face consistency, and appropriateness, acting as a human-in-the-loop mitigation.
[EXTERNAL_DOWNLOADS]: The documentation in references/channel-calibration.md references downloading video thumbnails from i.ytimg.com, which is an official Google service, for the purpose of performance analysis.

youtube-packaging