The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to upload local files (brand assets, source images) to a public hosting service, litterbox.catbox.moe, to convert them into URLs required by the Kie.ai generation tools. As a public host, any files uploaded are accessible to anyone with the resulting URL. While the skill instructs the agent to ask for permission before the first upload, this facilitates the movement of local data to an untrusted public service.
[COMMAND_EXECUTION]: The instructions direct the agent to execute shell scripts located within the skill's directory (host-reference.sh and save-generation.sh) to perform network operations and file management.
[DYNAMIC_EXECUTION]: The path to the execution scripts is resolved dynamically at runtime using a shell command that searches the user's home directory (ls -d "$HOME"/.claude/plugins/cache/...). This pattern introduces a risk of executing unintended scripts if the environment contains multiple plugin versions or is tampered with. Additionally, the skill uses python3 -c to execute dynamic code for housekeeping tasks.
[EXTERNAL_DOWNLOADS]: The save-generation.sh script downloads media assets from external URLs generated by the Kie.ai service to save them to the local filesystem.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data (URLs and generated files) and passes prompts to external AI models. 1. Ingestion points: Task result URLs from Kie.ai tools are passed to the save-generation.sh script. 2. Boundary markers: The skill includes an instruction to manually verify saved file content. 3. Capability inventory: The skill can write files and perform network uploads via shell scripts. 4. Sanitization: No sanitization of external URLs is described before they are passed to the scripts.

generate