The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation includes a manual installation command that fetches a shell script from the author's GitHub repository and pipes it directly to the shell.
Evidence in SKILL.md: curl -LsSf https://raw.githubusercontent.com/voidmatcha/ui-clone-skills/main/install.sh | bash
The skill explicitly instructs the agent not to auto-execute this command, instead prompting the user to run it manually if dependencies are missing.
[EXTERNAL_DOWNLOADS]: The skill is designed to navigate to and download data from arbitrary external websites as part of its primary capture function.
Ingestion occurs via the agent-browser tool for screenshots, video recording, and DOM metadata extraction.
[PROMPT_INJECTION]: The skill processes content from untrusted third-party websites, which constitutes an indirect prompt injection surface.
Ingestion points: agent-browser eval and screenshot commands in SKILL.md and detection.md read data from third-party sites.
Capability inventory: Includes file system operations (mkdir, cp), browser control (agent-browser), and video processing (ffmpeg).
Sanitization: SKILL.md and detection.md include explicit instructions for the agent to sanitize output, skip suspicious URIs, and redact directive-like text found in class names or attributes.
Boundary markers: The skill uses structured data (regions.json) to handle results, but no specific boundary markers are defined for the prompt interpolation itself.
[COMMAND_EXECUTION]: The skill executes various local commands to process captured assets and manage the local environment.
Tools used: agent-browser, ffmpeg, npx, mkdir, and mv.
Arguments for these commands are derived from user-provided URLs and CSS selectors detected during page analysis.

ui-capture