The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions facilitate the creation of interactive UI widgets that can trigger agent actions via the sendPrompt function, creating a surface for injection.
Ingestion points: The skill processes user-supplied data to generate visualizations, charts, and diagrams as defined in SKILL.md (e.g., "visualize data", "explain how X works").
Boundary markers: The instructions do not define boundary markers or safety delimiters to separate user data from the executable logic of the widgets.
Capability inventory: The skill utilizes the show_widget tool to execute HTML/JavaScript and the sendPrompt function to programmatically invoke the agent.
Sanitization: There is a lack of instructions for sanitizing or escaping user-controlled strings before they are interpolated into onclick handlers or <script> blocks.
[DYNAMIC_EXECUTION]: The skill's primary purpose involves generating and executing dynamic HTML, SVG, and JavaScript code fragments at runtime. The process of assembling these scripts through string concatenation introduces an inherent risk of code injection if user-provided content is included in the templates.
[METADATA_POISONING]: The skill metadata and documentation (in SKILL.md and README.md) claim to provide the "actual Anthropic 'Imagine — Visual Creation Suite' design rules" and the "complete Anthropic design system." These claims may be misleading regarding the official provenance of the guidelines, as the skill is provided by a third-party author.
[EXTERNAL_DOWNLOADS]: The skill fetches established libraries such as Chart.js and Mermaid.js from well-known content delivery networks, including cdnjs.cloudflare.com and esm.sh, to support its visualization features.

generative-ui