The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests user-supplied text descriptions to generate slides and schematics using remote LLMs. This creates a surface for indirect prompt injection where maliciously crafted slide content could attempt to influence the agent's output.
Ingestion points: User-provided prompts in scripts/generate_slide_image_ai.py and scripts/generate_schematic_ai.py.
Boundary markers: The scripts wrap user input with specific design guidelines and 'USER REQUEST' delimiters to minimize accidental instruction following.
Capability inventory: Uses subprocess calls to internal helper scripts and makes network requests via requests.post.
Sanitization: Relies on the security guardrails of the underlying LLM provider (e.g., Google Gemini models via OpenRouter).
[COMMAND_EXECUTION]: Several scripts use subprocess.run to orchestrate local tools and helper scripts.
Evidence: scripts/generate_schematic.py and scripts/generate_slide_image.py execute internal AI scripts using list-based argument passing, which avoids shell injection vulnerabilities. scripts/validate_presentation.py calls pdflatex with the -no-shell-escape flag enabled to prevent arbitrary command execution through LaTeX.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to external APIs to facilitate its core AI-powered features.
Evidence: scripts/generate_slide_image_ai.py and scripts/generate_schematic_ai.py perform POST requests to https://openrouter.ai/api/v1/chat/completions. OpenRouter is recognized as a well-known service aggregator for AI models.

scientific-slides