The Agent Skills Directory

[Indirect Prompt Injection] (HIGH): The skill is designed to process untrusted external content (the source material for the comic) and incorporate it into the generation workflow.
Ingestion points: User-provided content passed as a file argument (e.g., content.md) is ingested into the analysis and storyboard generation steps.
Boundary markers: The references/base-prompt.md uses a weak trailer instruction ("Please generate the comic page based on the content provided below:") without robust delimiters (like XML tags or random nonces) to encapsulate the untrusted data.
Capability inventory: The skill produces multi-page storyboards, character definitions, and detailed image prompts. A successful injection in the source content could force the agent to generate harmful imagery, exfiltrate the system prompt via the comic text, or ignore safety constraints in the resulting prompts.
Sanitization: There is no evidence of input validation, instruction-guarding, or escaping of the user content before it is processed by the LLM.

baoyu-comic