The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides detailed strategies and reference tables (specifically in references/moderation.md and references/styles.md) designed to bypass content moderation and copyright filters of AI image and video models. It includes a 'Safe-phrasing table' that instructs the agent to use specific synonyms and descriptive language (e.g., replacing 'bikini' or 'sexy' with 'ecommerce catalog photo' or 'confident') to avoid being flagged by moderation systems. It also describes how to replicate famous IP (e.g., 'The Simpsons', 'LEGO') using visual grammar to bypass brand-name filters.
[COMMAND_EXECUTION]: The skill's documentation in SKILL.md provides shell command templates for image manipulation using ffmpeg. These commands are used to crop character reference sheets into individual components (face and body close-ups), involving the execution of system tools with user-defined or generated file paths.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-controlled character premises and descriptions to construct prompts for external AI generation tools (SKILL.md). This creates an attack surface where instructions embedded in external user data could influence the behavior of the downstream tools.
Ingestion points: Character premises, descriptions, and attribute inputs provided by the user in the design process (e.g., Section 1 and 2 of SKILL.md).
Boundary markers: No explicit boundary markers or 'ignore embedded instructions' warnings are used in the prompt templates to separate user input from the system instructions.
Capability inventory: The skill uses Bash(curl), Bash(ffmpeg), and file Read/Write capabilities to generate and process media.
Sanitization: No validation or sanitization is performed on the user-provided character details before they are interpolated into the final tool calls.

varg-character-design