The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface.
Ingestion points: User-provided prompts are passed to generate.mjs and subsequently interpolated into system-like instructions for the model during the 'Prompt Rewriting' and 'Advanced Suggestions' phases.
Boundary markers: The script does not utilize explicit delimiters or safety warnings to separate the user-supplied prompt from the internal AI instructional templates.
Capability inventory: generate.mjs has the capability to read local files (fs.readFileSync), write files (fs.writeFileSync), and perform network requests to the Google Gemini API.
Sanitization: No sanitization or escaping is applied to the input prompt before it is sent to the API for processing.
[COMMAND_EXECUTION]: Local Script and Safety Configuration.
The skill executes a local Node.js script to orchestrate API calls. This script explicitly sets all HarmCategory thresholds to BLOCK_NONE to bypass API-side content filtering, a design choice intended to mirror the functionality of the reference application.
[DATA_EXFILTRATION]: File Access and Credential Management.
The skill accesses local image files via the --ref and --base flags, converting them to Base64 for transmission to the Google Gemini API.
Credentials (API keys) are managed via standard environment variables (GEMINI_API_KEY), following common security practices for CLI tools.
[EXTERNAL_DOWNLOADS]: Trusted Dependencies.
The setup instructions include the installation of @google/genai via npm. This is the official SDK provided by Google, which is a well-known and trusted service provider.

orugas-design-studio