The Agent Skills Directory

[PROMPT_INJECTION]: The prompt template in references/image.txt contains instructions designed to bypass the underlying AI model's safety and refusal mechanisms. It explicitly directs the model to generate "similar alternatives" for sensitive or copyrighted content and instructs it to "not refuse to generate the image."
[INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary external documents that are converted to markdown and subsequently used as input for story and image generation, creating a vulnerability surface where malicious instructions embedded in a document could influence the agent's behavior.
Ingestion points: The workflow in SKILL.md uses the markitdown tool to ingest user-provided input files in Step 1.
Boundary markers: The process lacks clear delimiters or protective instructions (e.g., "ignore embedded instructions") when interpolating the document content into the story and image generation prompts.
Capability inventory: The skill has the capability to execute shell scripts (scripts/gen_story.py, scripts/gen_image.py), perform network operations via API calls to Google Gemini and OpenAI, and write files to the local filesystem.
Sanitization: No sanitization or filtering is performed on the content extracted from the source document before it is processed by the AI models.

q-infographics