The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions in references/base-prompt.md that direct the downstream image generation model to bypass standard safety refusals regarding sensitive or copyrighted figures, instructing it to 'create stylistically similar alternatives' and 'DO NOT refuse to generate'.
[COMMAND_EXECUTION]: The workflow instructions in references/workflow/prompt-template.md and references/workflow/reference-images.md assume the agent has the capability to execute shell-like verification commands, such as test -f, to check for the existence of reference files before proceeding.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by ingesting untrusted data (article titles and summaries) and interpolating them into prompts for image generation tools without robust sanitization or delimiters.
Ingestion points: User-provided article content is saved to source.md and processed during the analysis step.
Boundary markers: The prompt template in references/workflow/prompt-template.md uses plain text headers (e.g., 'Article title:') to separate untrusted content from instructions, which may be insufficient to prevent content from influencing the agent's logic.
Capability inventory: The skill performs file system operations (read/write in refs/, prompts/, and config directories), executes status checks (implied test -f), and invokes external image generation tools.
Sanitization: No explicit sanitization or validation is defined for the content extracted from the source article before it is inserted into the final generation prompt.

baoyu-cover-image