The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/codex_ppt_runtime.py uses subprocess.run to execute system commands for managing the skill's execution environment.
It runs pip install -r requirements.txt during the bootstrap phase to install necessary dependencies.
It executes Python code strings via subprocess.run([python, "-c", code]) to verify that required libraries like openai, Pillow, and python-pptx are correctly installed.
[EXTERNAL_DOWNLOADS]: The skill downloads third-party Python packages from public registries to function.
requirements.txt specifies dependencies including python-pptx, Pillow, openai, and filelock.
The AtlasCloudImageProvider in scripts/image_providers/atlascloud.py makes external network requests to api.atlascloud.ai to generate and edit images.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (articles, papers, and reports) which is used to generate prompts for automated subagents.
Ingestion points: User-provided source material, outlines, and article summaries enter the system via outline.md and deck_spec.json (as described in SKILL.md).
Boundary markers: The skill attempts to mitigate injection by using structured JSON job files (prompts/slide_XX.json) to define isolated tasks for subagents.
Capability inventory: The skill has the capability to write files (origin_image/), execute local scripts (scripts/image_gen.py), and perform network operations for image generation.
Sanitization: There is no explicit sanitization or filtering of user-supplied text described before it is interpolated into the prompts used by the slide-worker subagents.
[DYNAMIC_EXECUTION]: The skill uses subprocess to verify the presence of libraries at runtime by passing string-based code to a Python interpreter, which is a form of dynamic execution.

codex-ppt