The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill retrieves visual references, stylesheets, and design patterns from a variety of public external sources including Kage, Minimal Gallery, Land-book, and Refero Styles.
[COMMAND_EXECUTION]: The skill employs the phoenix-ui CLI to interact with project contexts (phoenix-ui context) and link design briefs to specific project targets (phoenix-ui surface-brief link), ensuring design consistency across the repository.
[INDIRECT_PROMPT_INJECTION]: As the skill ingests untrusted data from external websites (HTML comments, metadata, and CSS), it is exposed to indirect injection. However, the instructions explicitly mandate that any commands or instructions discovered in scraped content must be treated as inert analysis material and never executed.
Ingestion points: External HTML, metadata, and 3rd-party prompts fetched via web scraping (SKILL.md).
Boundary markers: Explicitly instructs the agent to separate observed facts from source interpretations and to ignore embedded instructions (SKILL.md, Section 4).
Capability inventory: File writes to docs/design/, network fetches (http/https), and phoenix-ui CLI execution.
Sanitization: Prohibits passing scraped code to shell/Node and ignores requests to run commands or read credentials found in scraped content.
[DATA_EXFILTRATION]: The skill implements strict Server-Side Request Forgery (SSRF) protections. It specifically forbids access to localhost, internal network ranges (RFC1918), link-local addresses, and cloud metadata services. It also restricts the use of sensitive protocols like file: and data:, and prohibits the transmission of credentials within URLs.

design-reference