The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external media URLs (images, audio, video) provided by the user. \n
Ingestion points: Media URLs are passed into the runcomfy CLI via the --input JSON parameter in various routes described in SKILL.md. \n
Boundary markers: The CLI uses a structured JSON string for inputs, which helps separate data from the command structure, though the underlying models still process the untrusted media content. \n
Capability inventory: The skill utilizes the runcomfy CLI via Bash to execute model runs and writes output results to the local file system using the --output-dir flag. \n
Sanitization: The skill includes explicit safety warnings for the agent to suspect reference assets if behavior diverges from the prompt and provides guidelines for refusing harmful or non-consensual requests.\n- [EXTERNAL_DOWNLOADS]: The skill provides instructions to install the @runcomfy/cli package from the official NPM registry, which is a standard procedure for using the vendor's toolset.\n- [COMMAND_EXECUTION]: The skill performs command execution using the runcomfy CLI via Bash to interact with the RunComfy API. The scope of execution is restricted to the specific CLI tool as defined in the skill's allowed tools.

face-swap