The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides extractors that process untrusted data from external sources, presenting a surface for indirect injection attacks.
Ingestion points: The agent fetches and parses content from arbitrary website URLs (from-website.md), uploaded brand guide PDFs (from-pdf.md), and video keyframes (from-video.md) to generate design specifications.
Boundary markers: The instructions for extraction define the target fields but lack explicit warnings or delimiters to prevent the agent from following adversarial instructions that might be embedded in the source material.
Capability inventory: The skill generates visual-style.md files containing prompts for other AI models, produces HTML/CSS for slides, and integrates with external video generation platforms (HeyGen) via MCP tools.
Sanitization: There is no documented mechanism to filter or sanitize the extracted content before it is interpolated into fields like style_prompt_full, which are used to drive other tools.
[EXTERNAL_DOWNLOADS]: The extraction workflow involves fetching content from user-provided URLs. While this is a functional requirement for the skill, it involves interaction with arbitrary external domains to retrieve HTML and CSS data for analysis.
[DYNAMIC_EXECUTION]: The skill generates presentation content by creating HTML and CSS based on design system parameters (html-slides.md). This involves template-based generation of executable web code using data extracted from potentially untrusted external sources.

visual-style