The Agent Skills Directory

[PROMPT_INJECTION]: The skill's diagram generation utility in scripts/generate_schematic_ai.py is susceptible to indirect prompt injection. It accepts user-supplied descriptions and interpolates them directly into prompts sent to secondary AI models for review and refinement without using boundary markers or sanitization. This could allow maliciously crafted input to influence the AI model's output or potentially leak context.
Ingestion points: User-provided prompt argument in scripts/generate_schematic_ai.py.
Boundary markers: Absent in the review_prompt and improved_prompt templates.
Capability inventory: Subprocess execution (via wrapper) and file system writes (images and logs).
Sanitization: No input validation or filtering of the prompt text is performed before interpolation.
[COMMAND_EXECUTION]: The script scripts/generate_schematic.py uses subprocess.run to execute a localized helper script. While it processes user-provided text as an argument, it employs a list-based argument passing method (as opposed to a shell string), which prevents shell-level command injection and is considered a safe implementation of subprocess calls.
[EXTERNAL_DOWNLOADS]: The scripts/generate_schematic_ai.py script makes network requests to the OpenRouter API (openrouter.ai) to utilize remote AI models for schematic generation. Although this is a core intended feature, it requires an external API key and transmits user-defined content to a third-party service.

clinical-decision-support