The Agent Skills Directory

[PROMPT_INJECTION]: The skill features a URL and image analysis capability (hallmark study) that ingests untrusted data into the agent's context, creating a potential indirect prompt injection surface.
Ingestion points: Untrusted content from external URLs and screenshots is processed in references/study.md.
Boundary markers: Present; the instructions in references/study.md explicitly direct the agent to treat fetched content as "untrusted inert data" and ignore any commands found within it.
Capability inventory: The agent can write files to the local filesystem (e.g., design.md, tokens.css) and perform network requests using the WebFetch tool.
Sanitization: Sanitization is implemented via model instructions to ignore remote instructions and extract only inert design facts.
[EXTERNAL_DOWNLOADS]: The skill references a wide array of well-known third-party design services for icons, fonts, and placeholder imagery (e.g., Lucide, Phosphor Icons, Google Fonts, Unsplash). These references are legitimate and target established, safe domains.
[COMMAND_EXECUTION]: The skill directs the agent to utilize the WebFetch tool to retrieve HTML and CSS from user-provided URLs. The implementation includes detailed safety protocols to refuse local IP addresses, private network ranges, and non-web protocols, effectively mitigating common network-related risks.

hallmark