The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/capture_live_site.py uses subprocess.Popen to launch a headless browser (Chrome or Chromium). This is a central feature for the skill's purpose of capturing screenshots and computing CSS styles of web pages. The binary path is derived from the system PATH or the CAPTURE_CHROME environment variable. The command arguments are passed as a list, which prevents shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill is designed to fetch interaction patterns, accessibility criteria, and design documentation from various external sources, including the W3C (ARIA and WCAG), Google (Material Design), and Microsoft (Fluent Design). It also performs live-site captures. The skill implements several safeguards, such as validating robots.txt, enforcing crawl delays, requiring HTTPS, and maintaining a list of excluded hosts in the source-registry.yaml to prevent unauthorized scraping of subscription-based galleries.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by ingesting and processing untrusted data from web documentation and live-site renders. Malicious instructions could theoretically be embedded in the HTML or text of the surveyed sites.
Ingestion points: scripts/capture_live_site.py fetches external web pages, robots.txt, and terms of service.
Boundary markers: Not explicitly implemented in the script's output, but SKILL.md provides clear instructional boundaries for the agent to treat all fetched content as data rather than instructions.
Capability inventory: File system writes (to local evidence directories), network fetching via urllib.request, and local browser execution via subprocess.Popen.
Sanitization: SKILL.md explicitly mandates that all fetched documents must pass through a content-sanitization guardrail, with the result recorded in the vocabulary map.

visual-prior-art-survey