The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted image data using a Vision-Language Model (VLM). Images could contain adversarial text or patterns designed to override instructions.
Ingestion points: Local image files provided via the --input parameter in SKILL.md.
Boundary markers: Absent; the prompt templates do not use delimiters to isolate user instructions from the visual content of the image.
Capability inventory: The skill executes shell commands via uv run for image processing and media indexing.
Sanitization: No mechanisms are mentioned for sanitizing or filtering instructions that might be visually embedded in the input images.
[COMMAND_EXECUTION]: The skill executes local shell commands using uv run to invoke vendor-specific CLI tools, including comfy-imagedescribe and comfy-media, for generating descriptions and managing media galleries.

comfy-imagedescribe