The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input files (PPTX, images, text) to generate style packages containing natural language instructions for downstream AI models, creating a surface for indirect prompt injection.\n
Ingestion points: The skill ingests data from user-provided .pptx files via scripts/extract_pptx_style.py, images through the image-pipeline described in references/image-pipeline.md, and various markdown or text documents.\n
Boundary markers: It implements strong boundary controls by using a dedicated validation script scripts/validate_style_package.py that enforces a script-free, self-contained preview.html and validates all metadata against a fixed schema.\n
Capability inventory: The skill can write files to the local file system using scripts/write_style_package.py and execute subprocesses via scripts/capture_preview_webp.py.\n
Sanitization: Sanitization is handled by XXE (XML External Entity) protection in the PPTX parser, HTML escaping for content nodes in assets/preview.html.template, and magic bytes validation for image inputs.\n- [COMMAND_EXECUTION]: The skill executes external binaries to render style previews and convert images.\n
Evidence: The script scripts/capture_preview_webp.py uses subprocess.run to execute system browser binaries (Chrome, Chromium, or Edge) and the cwebp utility.\n
Context: These executions are restricted to rendering AI-generated preview HTML which is strictly validated for safety before execution. Commands are constructed as argument lists without using a shell environment, mitigating shell injection risks.

style-package-generate