The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user-supplied data (Salesforce code and image generation prompts) and incorporates it into requests sent to the Google Gemini model, creating a surface for indirect prompt injection.
Ingestion points: User code is pasted into templates in assets/review/apex-review.md and assets/review/lwc-review.md. Custom image prompts are provided as command-line arguments to scripts/generate_image.py.
Boundary markers: The review templates use basic text headers such as CODE: to separate input, providing minimal protection against adversarial instructions. Image prompts are interpolated directly into the API request.
Capability inventory: The skill utilizes network operations via the Google Gemini API, writes files to the local filesystem, and executes shell commands to open images.
Sanitization: There is no evidence of filtering or escaping user-provided content before it is processed by the external AI.
[EXTERNAL_DOWNLOADS]: The skill suggests a remote installation pattern for a prerequisite tool.
Evidence: The scripts/check-prerequisites.sh script recommends the command curl -LsSf https://astral.sh/uv/install.sh | sh to install the uv Python manager.
[COMMAND_EXECUTION]: The Python utility scripts/generate_image.py executes a system command to display output files to the user.
Evidence: The script uses subprocess.run(["open", output_path]) to launch the default image viewer on macOS.

sf-diagram-nanobananapro