The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a centralized management script (scripts/codex_ppt_runtime.py) to handle its execution environment. It uses subprocess.run to create virtual environments, install necessary dependencies via pip, and execute diagnostic checks (e.g., verifying imports). These commands are restricted to the skill's primary purpose of maintaining its own local environment.
[EXTERNAL_DOWNLOADS]: The skill performs downloads from the official Python Package Index (PyPI) to install dependencies such as python-pptx, Pillow, and the openai SDK. These are well-known, version-pinned libraries required for generating slides and interacting with image APIs.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (articles, papers, or notes) to generate presentation slides. This creates a structural risk for indirect prompt injection.
Ingestion points: Untrusted data is read from source files and stored in orchestration files like outline.md and deck_spec.json.
Boundary markers: The skill uses a structured JSON-based orchestration system (prompts/slide_XX.json) and clear Markdown section headers to delimit instruction blocks from content in the subagent prompts.
Capability inventory: The skill has broad capabilities including local file system access (writing PPTX and image files) and network API access (sending data to OpenAI or AtlasCloud APIs).
Sanitization: The scripts do not perform explicit sanitization or filtering of the source text, relying on the model's ability to interpret specific schema fields.

codex-ppt