The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on a local Python script (scripts/avs.py) to perform its core routing and search logic. It also instructs the agent to use ffprobe for technical metadata inspection of media files.
[INDIRECT_PROMPT_INJECTION]: The skill manages a catalog of 193 'upstream' skill paths. This creates an attack surface where untrusted data (the content of those upstream skills) enters the agent's context.
Ingestion points: Information is ingested from references/registry/retrieval.jsonl and external URLs linked in references/catalog.md.
Boundary markers: Present. The SKILL.md and various guide files (e.g., references/models/wan.md, references/quality.md) explicitly instruct the agent to treat external content as untrusted source material and never execute unreviewed commands.
Capability inventory: python3 scripts/avs.py (logic and routing) and ffprobe (metadata extraction).
Sanitization: Relies on instructional guardrails and explicit 'preflight' review steps rather than automated filtering.
[INDIRECT_PROMPT_INJECTION]: The skill uses a structured workflow system (references/registry/workflows.json) that defines sequences of actions for the agent to take. These instructions influence agent behavior in subsequent turns of the conversation.

video-router