The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various external binaries via shell commands, including soffice (LibreOffice), pdftoppm, and git for document processing and validation tasks.
Evidence found in scripts/thumbnail.py, scripts/office/soffice.py, and scripts/office/validators/redlining.py.
[REMOTE_CODE_EXECUTION]: The skill utilizes a dynamic execution pattern by compiling C code at runtime and injecting it into a target process.
In scripts/office/soffice.py, a C shim is written to a temporary file, compiled with gcc, and loaded into the soffice process via the LD_PRELOAD environment variable to modify its socket handling behavior.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests and extracts data from untrusted PowerPoint presentations.
Ingestion points: Slide XML files (ppt/slides/*.xml) and layout files are read during the editing and analysis workflows described in SKILL.md and editing.md.
Boundary markers: No delimiters or instructions to ignore nested prompts are used when the agent processes extracted text.
Capability inventory: The skill possesses capabilities for command execution (scripts/thumbnail.py), runtime code compilation (scripts/office/soffice.py), and file system manipulation (scripts/office/pack.py).
Sanitization: Natural language content is not sanitized after extraction; however, defusedxml is correctly used to protect against XML-based attacks like XXE.
[EXTERNAL_DOWNLOADS]: The skill documentation instructs the installation of several well-known Python and Node.js packages from established registries.
Fetches markitdown, Pillow, and pptxgenjs from official package registries.

pptx