The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it downloads and parses metadata (titles, descriptions, tags) from external video platforms (via scripts/video_downloader.py) and feeds this untrusted data into its "Explosion Analysis" agent. An attacker could potentially craft video metadata that attempts to manipulate the agent's output during the script creation phase.
Ingestion points: Video titles, descriptions, and tags extracted from user-supplied URLs in scripts/video_downloader.py and processed in references/agents-prompts.md.
Boundary markers: The prompt instructions lack explicit boundary markers or warnings to the agent to disregard instructions embedded in the analyzed video data.
Capability inventory: The skill has capabilities for file system writes (./downloads), network operations (Coze API interactions), and multi-agent coordination.
Sanitization: There is no evidence of sanitization or filtering of the external metadata before it is interpolated into the agent's context.
[EXTERNAL_DOWNLOADS]: The skill uses the standard yt-dlp library to fetch media and metadata from external sources. This is a core part of the skill's functionality for analyzing viral content and is considered a legitimate use of network resources.
[COMMAND_EXECUTION]: The skill involves executing specific local scripts (scripts/video_downloader.py and scripts/coze_video_api.py) to perform its tasks. These executions are constrained to the skill's own internal tools and follow best practices for secret management by using environment variables for API credentials.

pet-commerce-creator