The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted inputs (academic papers or text documents) to automatically generate slide deck outlines, prompts, and content. If a document contains adversarial text patterns, it could hijack the prompt generation pipeline or influence the image generator model.
Ingestion Point: SKILL.md Workflow Step 1 parses external PDFs or markdown files (source.md).
Boundary Markers: Missing or inadequate structure to segregate untrusted content from system instructions when assembling files in the prompts/ directory.
Capability Inventory: The skill executes multiple local TypeScript/Python scripts, modifies files, and performs external API calls to Gemini.
Sanitization: The code does not perform active sanitization or filter out prospective instructional overrides from the processed paper text.
[COMMAND_EXECUTION]: scripts/generate-slides.py calls subprocess.check_call() to invoke pip install. This execution is safe and bounded as it relies on a hardcoded array configuration without injecting dynamic user arguments.
[EXTERNAL_DOWNLOADS]: The Python script dynamically ensures the presence of the google-genai package from the official PyPI repository at runtime. Since this points directly to a well-known package provider, it presents no immediate safety concerns.

paper-slide-deck