The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (image URLs) which are subject to indirect prompt injection (image-based).
Ingestion points: Untrusted data enters via the image URL field and the prompt field defined in the model schema in SKILL.md.
Boundary markers: The documentation explicitly acknowledges image-based prompt injection as a known risk for image-edit models.
Capability inventory: The skill uses the runcomfy CLI to execute image processing, involving network operations and file writes in SKILL.md.
Sanitization: The CLI transmits prompts as JSON strings to avoid shell expansion and enforces a 2 GiB download cap to prevent disk-exhaustion attacks.
[EXTERNAL_DOWNLOADS]: The skill guides the user to install the official @runcomfy/cli utility from the npm registry and fetches model outputs from the RunComfy infrastructure.
[COMMAND_EXECUTION]: The skill invokes the runcomfy CLI tool to perform model runs, login operations, and result downloads.

flux-kontext