The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze academic papers (PDFs) and extract data to generate presentation outlines, speaker scripts, and Q&A pairs. This process creates a surface for indirect prompt injection where instructions embedded in a research paper could attempt to manipulate the agent's behavior.
Ingestion points: Research papers are ingested as PDF files in SKILL.md (Phase 0) and processed by extract_pdf_figures.py and check_text_overflow.py.
Boundary markers: The skill does not implement explicit delimiters or instructions to the model to ignore formatting or commands embedded within the source PDF text.
Capability inventory: The skill possesses the ability to write files, execute shell commands via Bash, and generate Python scripts (SKILL.md, build_pptx_nature_lancet.py).
Sanitization: There is no evidence of filtering or sanitization of the text extracted from the papers before it is interpolated into model prompts or generated scripts.
[DYNAMIC_EXECUTION]: As part of its Phase 3 workflow, the skill generates a presentation-specific Python script (inject_notes.py) containing speaker notes parsed from the paper.
This generated script is intended to be executed to modify the PowerPoint files. If the content from the paper (e.g., specific quotes or data points) is not strictly escaped by the agent during the generation of this script, it could lead to code injection within the local environment.
[COMMAND_EXECUTION]: Several helper scripts (check_text_overflow.py, extract_pdf_figures.py, trim_caption.py) execute external system utilities to process documents.
The skill calls pdftotext, pdftoppm, and soffice (LibreOffice) using subprocess.run. While these calls use the safe list-based format, they operate on file paths derived from user input or extracted paper data.

present-paper