The Agent Skills Directory

[DYNAMIC_EXECUTION]: Evidence in scripts/office/soffice.py shows that the skill generates C source code at runtime to shim AF_UNIX socket calls. This code is written to a temporary file, compiled into a shared library using gcc, and then loaded into the LibreOffice environment via the LD_PRELOAD environment variable to modify process behavior.
[COMMAND_EXECUTION]: Several components use subprocess.run to execute system-level commands, including gcc for the runtime shim compilation, git for document redlining comparisons in redlining.py, and Office automation tools like soffice and pdftoppm in thumbnail.py.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from PowerPoint and Word documents, which can contain hidden instructions intended to override agent behavior.
Ingestion points: Document content is ingested via unpack.py and external parsing tools.
Boundary markers: XML content and slide text are extracted into the agent's context without clear delimiters or instructions to ignore embedded commands.
Capability inventory: The skill possesses powerful capabilities including file writing, shell command execution, and runtime code compilation.
Sanitization: The skill uses defusedxml to prevent XML-based attacks (XXE), but does not sanitize the natural language content extracted from documents.
[METADATA_POISONING]: The LICENSE.txt file identifies Anthropic, PBC as the copyright holder, which conflicts with the provided author name 'bitcjm'. This discrepancy in metadata could be deceptive regarding the skill's official status or origin.

pptx