The Agent Skills Directory

[OBFUSCATION]: The file demo/index.html contains a large Base64-encoded string (encodedHtml) representing an entire HTML design demo. This encoding obscures the content from static analysis tools.
[DYNAMIC_EXECUTION]: The skill uses JavaScript in demo/index.html to decode the encodedHtml string at runtime and inject it into an iframe using srcdoc. It also dynamically converts Base64 asset data into Blobs using URL.createObjectURL, which involves runtime assembly and execution of content.
[INDIRECT_PROMPT_INJECTION]: The demo/PROMPT.md file contains instructions for the agent to visit and analyze an external URL (https://neuform.ai/...) to match its design fidelity. This exposes the agent to potential malicious instructions hidden in the remote content.
Ingestion points: The agent is directed to visit the design URL via demo/PROMPT.md.
Boundary markers: Absent; no instructions are provided to ignore embedded commands.
Capability inventory: The agent is tasked with HTML/CSS code generation; no sensitive system tools are directly invoked.
Sanitization: Absent; external content is used directly as a reference for code generation.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill's demo configuration in demo/source.json and the decoded HTML reference external runtime libraries and assets from sources like Google Fonts, Cloudflare (GSAP), and Supabase. These are well-known services and the references are consistent with the skill's purpose.

container-lines