The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill explicitly instructs the agent to "下载并使用需要的任何字体使其成为现实" (Download and use any fonts needed to make it a reality). This encourages the agent to perform network operations to fetch binary assets from unverified external sources.
[COMMAND_EXECUTION]: The instructions direct the agent to "搜索 ./canvas-fonts 目录" (Search ./canvas-fonts directory). This requires file system enumeration and tool invocation to process font assets located in the local environment.
[PROMPT_INJECTION]: The "最终步骤" (Final Steps) section utilizes a simulated user quote ("用户已说「它不够完美...」") to trigger a self-refinement loop. This is an instruction-based behavioral override designed to force the agent into iterative processing without actual user input.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests user-provided "subtle concept clues" and instructions to generate complex visual documents.
Ingestion points: User prompts for posters, artwork, and design projects.
Boundary markers: Absent; there are no delimiters defined to separate untrusted user references from the skill's operational logic.
Capability inventory: File system read/write (.md, .pdf, .png), directory traversal (./canvas-fonts), and network operations for font downloads.
Sanitization: Absent; the skill does not specify any validation or filtering of user-provided conceptual inputs before processing them.

canvas-design