The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill incorporates untrusted user input directly into instructions for a third-party image generation model, creating a surface for indirect prompt injection where malicious text could influence the model's behavior.
Ingestion points: Variables such as {书名} (book title), {作者名} (author name), and {REF_IMAGE} (reference image URL) are sourced from the user and interpolated into the generation prompt.
Boundary markers: The skill does not employ delimiters or explicit instructions for the AI to ignore potential injection patterns within these user-provided strings.
Capability inventory: The skill utilizes powerful tools including curl for network access, magick/sips for image manipulation, and standard shell utilities (jq, base64) for data processing.
Sanitization: No sanitization, escaping, or validation logic is present for the user-supplied strings before they are included in the prompt construction.
[EXTERNAL_DOWNLOADS]: The skill performs automated downloads from external, user-specified locations.
Evidence: The script provided in SKILL.md (Step 4, 图生图) uses curl -fsSL to download image files from a dynamic $REF_IMAGE URL provided by the user into a temporary local directory.

story-cover