The Agent Skills Directory

[DYNAMIC_EXECUTION]: The workflow for creating presentations from scratch involves generating individual JavaScript files for each slide and then loading them at runtime via require() within a master compilation script (compile.js). While the file paths are numerically constrained, this pattern of dynamic loading and execution of generated code is a significant capability.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data by reading and editing existing PowerPoint files through text extraction (markitdown) and XML manipulation. This creates an attack surface where maliciously crafted files could attempt to influence the agent's context or behavior.
Ingestion points: User-provided PPTX files processed via markitdown and Python zipfile extraction (references/editing.md).
Boundary markers: None implemented in the provided instructions.
Capability inventory: Subprocess execution for compilation and extraction, filesystem read/write, and dynamic module loading.
Sanitization: The skill recommends using defusedxml to prevent XML vulnerabilities, but does not explicitly instruct the agent to validate file paths during the ZIP extraction process, potentially exposing it to path traversal risks.
[DATA_EXPOSURE_AND_EXFILTRATION]: The presentation editing workflow involves unpacking PPTX archives (which are ZIP files) into a /tmp/ directory. Without unique, restricted directory paths and explicit path validation, this process could be vulnerable to archive-based attacks (ZipSlip) if processing a malicious file.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill requires the installation of several external packages from public registries (NPM and PyPI), including pptxgenjs, markitdown, sharp, and react-icons. While these are well-known libraries, they involve fetching and executing code from external repositories during environment setup.

pptx-generator