The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process untrusted data from users and external sources, which creates a surface for indirect prompt injection attacks.
Ingestion points: The skill takes user-provided task descriptions as input for the vg generate models command. Furthermore, the modality reference files (e.g., audio-to-text.md, image-to-image.md) document the use of audio_url, image_url, prompt, and query parameters which are passed to external AI models.
Boundary markers: The documentation lacks instructions for the agent to use delimiters or sanitization techniques to distinguish between system instructions and data originating from these untrusted inputs.
Capability inventory: The skill relies on the vg generate CLI for network communication with remote AI providers and supports file system writes via the --download flag in multiple processing chains (e.g., in references/image-to-3d.md).
Sanitization: No evidence of input validation or sanitization is present in the provided instructions.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of a global Node.js package from a non-standard source.
Evidence: SKILL.md instructs the agent to run npm install -g vibedgames to enable its functionality.
[COMMAND_EXECUTION]: The instructions frequently use shell command execution to perform core tasks, including model discovery, schema inspection, and executing media generation workflows.
Evidence: Examples like vg generate run, vg generate models, and vg generate upload are used extensively throughout the skill and its references.

model-catalog