The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions in references/extractors/from-website.md, references/extractors/from-pdf.md, and references/extractors/from-video.md direct the agent to fetch and analyze content from external URLs provided by the user. While this is a primary function of the skill, it involves interacting with potentially untrusted remote content.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its processing of untrusted external data to generate a style_prompt_full field, which is later used as instructional input for other tools (e.g., HeyGen, Figma).
Ingestion points: External content is ingested via references/extractors/from-website.md (website URLs), references/extractors/from-pdf.md (uploaded PDFs), and references/extractors/from-video.md (video URLs).
Boundary markers: Analysis of the extraction prompts in the referenced files shows an absence of explicit boundary markers or instructions to ignore embedded natural language commands within the source data.
Capability inventory: The skill generates executable HTML/CSS code in references/connectors/html-slides.md and instructs the agent to call external tools (HeyGen Video Agent) in references/connectors/heygen-video-agent.md using the extracted style prompts.
Sanitization: There is no evidence of sanitization or validation logic to filter out potentially malicious instructions from the extracted visual style fields before they are passed to downstream connectors.

visual-style