The Agent Skills Directory

[COMMAND_EXECUTION]: The script gemini_mockup.py executes system-level commands using subprocess.run to trigger google-chrome or chromium-browser for rendering HTML mockups. Additionally, run.sh utilizes fuser and xdg-open for process management and opening browser instances.
[DATA_EXFILTRATION]: The gemini_mockup.py script bundles sensitive local content, including source code files, project screenshots, and design tokens, into a ZIP archive which is then sent to the Gemini API (generativelanguage.googleapis.com). This creates a significant data exposure surface for the local workspace.
[EXTERNAL_DOWNLOADS]: The skill uses the httpx library in gemini_mockup.py to fetch external content from arbitrary URLs provided as design references. It also performs runtime installation of the @google/stitch-sdk via npm install and references unverified dependencies from local experimental directories.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. Ingestion points: Design specs, image files, and external URLs are ingested via CLI arguments in run.sh and gemini_mockup.py. Boundary markers: The skill lacks delimiters or instructions to prevent the agent from obeying commands embedded within ingested design references. Capability inventory: The skill has access to the local file system, can perform network requests, and execute shell commands. Sanitization: Ingested data is passed to LLM endpoints without sanitization or validation.

mockup-lab