The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the RunComfy service by executing the runcomfy CLI. It securely passes user instructions and configuration as JSON strings, which prevents traditional shell injection attacks.
[EXTERNAL_DOWNLOADS]: The skill requires the @runcomfy/cli tool and downloads generated images from the vendor's verified domains (runcomfy.net and runcomfy.com). It includes safety measures such as a 2 GiB file size limit to protect local disk space from malicious or runaway outputs.
[CREDENTIALS_UNSAFE]: The skill manages authentication using a dedicated configuration file at ~/.config/runcomfy/token.json or through an environment variable. This is a standard and secure method for CLI tools to manage access without exposing system-wide credentials.
[PROMPT_INJECTION]: The skill processes user-supplied prompts and external image URLs, which is a known surface for indirect prompt injection.
Ingestion points: User-provided strings in the prompt field and external HTTPS image URLs in the images array (SKILL.md).
Boundary markers: User inputs are encapsulated within structured JSON objects before being passed to the CLI.
Capability inventory: The skill executes the runcomfy CLI and downloads files to the user's local filesystem (SKILL.md).
Sanitization: The skill documents that the CLI transmits data directly to the Model API over HTTPS without shell expansion, and it explicitly warns users about the inherent risks of image-based prompt injection.

gpt-image-2