The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, including reusable style fragments in generations/styles.json and user-supplied reference images in generations/refs/, which are then used to build prompts for AI generation.
Ingestion points: SKILL.md describes loading prompt fragments from generations/styles.json; recipe files in the models/ directory (e.g., models/gpt-image-2.md) use reference image paths.
Boundary markers: The instructions do not provide delimiters or specific 'ignore embedded instructions' warnings when these inputs are interpolated into model calls.
Capability inventory: The skill has the ability to execute shell commands via kie-cli (subprocess calls) and perform file-system writes to the generations/ directory.
Sanitization: No sanitization or validation logic is defined for the content of style fragments or image metadata.
[COMMAND_EXECUTION]: The model recipe files (such as models/nano-banana-2-lite.md) instruct the agent to execute shell commands using kie-cli. These commands include arguments like --prompt and --refs which are directly derived from user input and external file content.
[EXTERNAL_DOWNLOADS]: For asynchronous video generation models (e.g., models/kling-3.0.md, models/veo-3.1.md), the skill's defined workflow involves downloading media files from external URLs provided by the AI service provider (Kie AI) after the task is completed.

generate