The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to clone several GitHub repositories into the ComfyUI/custom_nodes/ directory, including ComfyUI-Manager, rgthree-comfy, ComfyUI-KJNodes, and ComfyUI_essentials. While these are widely used in the ComfyUI ecosystem, they are third-party extensions maintained by individual community members.
[REMOTE_CODE_EXECUTION]: By cloning these custom node repositories, the agent is instructed to incorporate external Python code into the ComfyUI environment, which is executed at runtime to facilitate the workflow.
[EXTERNAL_DOWNLOADS]: The skill requires downloading approximately 39GB of model weights and VAE files from Hugging Face. Sources include the Comfy-Org/Ideogram-4 (a well-known organization) and Aitrepreneur/FLX (a third-party mirror).
[INDIRECT_PROMPT_INJECTION]: The skill is configured to process files matching the **/*.json glob, which are used as structured captions for image generation. This configuration creates a vulnerability surface for indirect instructions embedded in user-provided or externally sourced JSON data.
Ingestion points: Processes JSON files containing high_level_description and compositional_deconstruction fields via the CLIPTextEncode and Ideogram4PromptBuilderKJ nodes.
Boundary markers: No specific boundary markers or 'ignore' instructions are provided in the workflow description or JSON templates to prevent the model from following instructions found inside the data.
Capability inventory: The workflow has the capability to generate images, apply post-processing sharpening, and save files to the local disk using the SaveImage node.
Sanitization: No sanitization or validation logic is specified for the input JSON strings, which are directly interpolated into the CLIP text encoder.

ideogram-ultra