The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user-provided URLs and local configuration files to populate design variations. Evidence Chain: 1. Ingestion points: The skill uses WebFetch for live URLs (references/ingestion-cascade.md) and repository scanning (SKILL.md). 2. Boundary markers: Scraped data is stored in a structured brief JSON object within an HTML comment. 3. Capability inventory: Capabilities include network fetching, file reading/scanning, and artifact generation. 4. Sanitization: The skill implements a 'hard gate' sanitization process that escapes all special characters for text-node rendering and validates colors, fonts, and URLs against strict allowlists.
[EXTERNAL_DOWNLOADS]: The skill utilizes WebFetch to retrieve content from external websites for the purpose of UI design extraction. This is protected by hardening rules such as a 10-second timeout, maximum of 3 redirects, and strict content-type enforcement (HTML/XHTML only) with no JavaScript execution.
[COMMAND_EXECUTION]: The skill employs local file scanning (Grep, Glob) to extract design system tokens and marketing copy from project files. These operations are limited to static analysis of configuration and markdown files.
[DYNAMIC_EXECUTION]: The skill generates and renders HTML/CSS variations as a primary feature. It enforces security gates to prevent the inclusion of executable scripts, inline event handlers, or data-uri based script injection in the final produced artifacts.

design-variations