The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install an external CLI tool via the vibedgames npm package. This package originates from an unknown source not included in the trusted vendors or well-known services lists.
[DATA_EXFILTRATION]: The workflow involves uploading local media files to an external server (vibedgames) for processing. While this is central to the skill's utility, the data is transmitted to a non-whitelisted and unverified domain.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (such as product images, brand guides, and character references) and incorporates it into instructions for AI models and tool parameters.
Ingestion points: User-provided source assets, brand guides, and transcript files identified in SKILL.md and references/commercial.md.
Boundary markers: No explicit delimiting or 'ignore' instructions are present to isolate untrusted data from the execution context.
Capability inventory: The skill performs shell command execution (vg generate run), file uploads (vg generate upload), and local file downloads (vg generate status --download) across multiple recipes.
Sanitization: The skill employs a dedicated 'Extractor' node designed to convert planner output into deterministic, low-creativity prompts as a defensive measure, as documented in references/node-rules.md.
[COMMAND_EXECUTION]: The skill executes shell commands via the vg CLI tool using parameters derived from processed media and user descriptions. These commands may be influenced by untrusted inputs during the media production lifecycle.

media-workflow