The Agent Skills Directory

[OBFUSCATION]: The demo implementation in demo/index.html stores its primary source code as an encoded string.
Evidence: The encodedHtml constant contains a large Base64-encoded string representing the complete HTML, CSS, and JavaScript of the design demo.
[DYNAMIC_EXECUTION]: The skill's demo environment dynamically decodes and executes content at runtime.
Evidence: The script in demo/index.html uses atob and TextDecoder to process the encoded HTML and injects it into an iframe via the srcdoc property.
[EXTERNAL_DOWNLOADS]: The demo fetches runtime libraries and assets from well-known external services and cloud providers.
Evidence: demo/source.json and the decoded HTML reveal dependencies on Cloudflare's CDN (GSAP, ScrollTrigger), Google Fonts, Fontshare, and asset storage on Supabase.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external brand data and content, creating an attack surface for embedded instructions.
Ingestion points: The remix prompt in demo/PROMPT.md instructs the agent to ingest external brand, subject, and copy details.
Boundary markers: No explicit instruction delimiters or sanitization warnings are present to isolate user-supplied content from the skill's logic.
Capability inventory: While primarily a design tool, it specifically targets "login sections" and authentication flows, which could be exploited for phishing if malicious inputs are processed.
Sanitization: There is no evidence of input validation or content filtering.
[METADATA_POISONING]: The skill uses simulated social proof and ranking data within its prompts to establish authority.
Evidence: demo/PROMPT.md and demo/source.json include specific view counts, favorite counts, and ranking positions (#1) designed to make the skill appear highly popular and verified.

split-layout-technical