The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external files which could contain malicious instructions designed to influence subsequent automated workflows.
Ingestion points: ../../references/REFERENCE_SHEETS.md and reference_sheet_plans.json (referenced in SKILL.md).
Boundary markers: None specified in the instructions.
Capability inventory: Execution of ComfyUI workflows and dependency validation.
Sanitization: The instructions mention "validation" but do not define specific sanitization or escaping mechanisms for the ingested content.
[DYNAMIC_EXECUTION]: The skill executes workflows stored in the comfyui_workflows/ directory. The documentation explicitly states these are "real editable workflow sources, not sanitized blueprints," and mentions validating "live dependencies" before execution, implying potential runtime changes or external code loading.
[COMMAND_EXECUTION]: The skill uses relative path traversal (../../references/REFERENCE_SHEETS.md) to access files outside its own directory structure. While potentially legitimate for project-integrated tools, this pattern allows the agent to read arbitrary files if the environment is not properly sandboxed.

reference-sheets