The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references the installation of the 'belt' CLI and related skills from the vendor's own repositories (e.g., belt-sh/cli and inference-sh/skills). These are resources owned by the skill author ('inference-sh').
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: User-provided text prompts are ingested through the belt app run command's --input parameter in SKILL.md.
Boundary markers: The prompts are encapsulated within structured JSON objects, which helps delimit data from command parameters.
Capability inventory: The skill is restricted to using the belt CLI tool within the Bash environment (allowed-tools: Bash(belt *)), limiting the potential impact of any injected instructions.
Sanitization: Input validation and sanitization are handled by the external belt CLI and the backend API service rather than the skill instructions themselves.

google-veo