The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/office.py script executes the soffice (LibreOffice) binary via subprocess.run to convert PowerPoint documents to PDF and images. The command is constructed using a list of arguments, which mitigates shell injection risks, and it utilizes a temporary user profile (-env:UserInstallation) to ensure the execution environment is isolated from the host user's configuration.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted PowerPoint files which may contain embedded instructions.
Ingestion points: The agent reads and evaluates slide text, speaker notes, and shape metadata using python-pptx and the provided utility scripts as described in the SKILL.md workflow steps 1, 2, and 4.
Boundary markers: There are no explicit instructions or delimiters provided to the agent to isolate untrusted presentation data from its own instructional context.
Capability inventory: The agent can write to the local file system, execute local Python scripts, and invoke the LibreOffice binary.
Sanitization: While the scripts use defusedxml to mitigate XML-based attacks (XXE) and perform ZIP path validation, the high-level semantic content is not sanitized against adversarial prompt injection.
[EXTERNAL_DOWNLOADS]: The skill uses uv run to dynamically install well-known Python packages from official registries, including python-pptx, defusedxml, pypdfium2, pillow, and lxml to support its core functionality.

pptx