The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from user-provided design briefs and external references (URLs and images) to inform its design decisions. This constitutes a potential attack surface for indirect prompt injection.
Ingestion points: Processes user-provided text briefs and reference signals such as linked URLs and screenshots (SKILL.md, Section 0.A).
Boundary markers: Implements a mandatory "Design Read" step (Section 0.B) requiring the agent to summarize its interpretation of the brief before generating any code, which acts as a manual verification layer.
Capability inventory: The agent has the capability to read project configuration files (e.g., package.json) and generate frontend source code (SKILL.md, Section 3 and 11).
Sanitization: While no explicit text sanitization is defined, the skill enforces a rigid "Pre-Flight Check" (Section 14) that heavily constrains the agent's output space, reducing the risk of malicious payload execution.
[EXTERNAL_DOWNLOADS]: The skill recommends the installation and use of official design systems and assets from well-known services and trusted organizations, including Google, Microsoft, IBM, GitHub, and Shopify (SKILL.md, Section 2.A and Appendix A).
Well-known services: Uses picsum.photos for placeholder imagery and cdn.simpleicons.org for brand logos.
[COMMAND_EXECUTION]: Instructs the agent to perform standard development environment operations, such as checking package.json for dependencies and using package managers (npm, npx, yarn) to install legitimate UI libraries (SKILL.md, Section 3.F and Appendix A).

taste-skill