The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the runcomfy CLI to process video generation requests. It passes configuration as a structured JSON object to the --input flag, which mitigates the risk of shell injection by ensuring prompt text and URLs are treated as data rather than executable shell code.
[EXTERNAL_DOWNLOADS]: The skill requires the @runcomfy/cli package, which is the official tool provided by the vendor for interacting with their API. It also references standard installation procedures for the skill itself via the genmedia-labs repository on the skills platform.
[INDIRECT_PROMPT_INJECTION]: The skill identifies a potential attack surface where instructions could be embedded in reference media processed by the agent. It documents the following evidence chain:
Ingestion points: External media files (images, videos, and audio) are ingested via URLs provided in the images, videos, and audios fields of the input schema (SKILL.md).
Boundary markers: The instructions explicitly mandate that the agent treat any text visible or audible inside reference media as data and never as instructions, while only using reference URLs explicitly supplied by the user for the current generation.
Capability inventory: The skill uses the runcomfy CLI tool (via allowed-tools: Bash(runcomfy *)) to execute remote generations and save output to the local filesystem.
Sanitization: The instructions provide behavioral sanitization by requiring the agent to disregard any commands, watermark text, or overlays found within the media content to prevent acting on adversarial instructions.
[CREDENTIALS_SAFE]: The skill correctly handles API tokens by identifying the local storage path (~/.config/runcomfy/token.json) and recommending safe practices, such as using environment variables (RUNCOMFY_TOKEN) for CI environments and preventing the logging or echoing of secrets into the conversation context.

seedance-2-5-reference-to-video