The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes research papers which are external, untrusted data sources, creating a potential surface for indirect prompt injection.
Ingestion points: The skill reads compiled papers in PDF or LaTeX format to extract key insights, contributions, and architecture details (found in SKILL.md under Workflow Step 1).
Boundary markers: The instructions lack specific guidance on using delimiters or system-level warnings to ensure the agent ignores any malicious instructions that might be embedded within the source paper's text.
Capability inventory: The skill possesses capabilities to write files to the local file system (prs.save() in the Python script) and execute shell commands to compile the presentation (latexmk -pdf, python3 generate_slides.py).
Sanitization: There is no evidence of sanitization or filtering of the text extracted from the paper before it is interpolated into the Beamer LaTeX source or the Python generation script.
[DYNAMIC_EXECUTION]: The skill generates and executes local scripts and LaTeX source code to produce its final outputs.
Evidence: The skill provides a complete Python script (generate_slides.py in references/slide-templates.md) and a LaTeX Beamer template, which are meant to be populated with paper content and then executed using the system's Python interpreter and LaTeX compiler.

presenting-conference-talks