The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the extraction of design systems from external, untrusted sources including website URLs, PDF brand guides, and video keyframes. The extracted data is used to populate fields such as style_prompt_full, which is then applied as an instruction prompt for other AI tools (e.g., HeyGen Video Agent) or as design specifications for Figma and paper.design. This workflow creates a vulnerability surface where a malicious website or document could contain hidden instructions designed to influence the agent's extraction behavior or poison the resulting design prompt for downstream tools.
Ingestion points: Defined in references/extractors/from-website.md, references/extractors/from-pdf.md, and references/extractors/from-video.md where the agent is instructed to fetch and analyze remote content.
Boundary markers: The instructions do not specify the use of delimiters or 'ignore embedded instructions' warnings when processing extracted content.
Capability inventory: The skill produces high-level generative prompts for video agents, Figma design instructions, and raw HTML/JS code for presentations.
Sanitization: There is no mention of sanitizing or validating the text extracted from external sources before it is interpolated into new prompts or code.
[EXTERNAL_DOWNLOADS]: The skill references external resources and frameworks for its implementation logic.
Evidence: The file references/connectors/html-slides.md references the zarazhangrui/frontend-slides GitHub repository as a target for generated HTML. Additionally, the design format specification in references/spec.md mandates that all assets (images, icons, snippets) be hosted externally and referenced via URL, creating a dependency on external, user-provided infrastructure.

visual-style