The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a vulnerability surface for indirect prompt injection as it processes untrusted data from external sources (blog URLs, GitHub PR diffs, meeting notes) to generate scripts for video production.
Ingestion points: External data enters the context through web_fetch, gh pr view, and direct text pasting as documented in SKILL.md.
Boundary markers: The skill does not implement delimiters or explicit instructions for the agent to ignore potentially malicious instructions embedded in the ingested content.
Capability inventory: The skill performs file system writes (mp4, srt, images) and executes shell commands via subprocess.run to orchestrate ffmpeg and AI model inference.
Sanitization: No explicit sanitization or filtering logic is present in the scripts before external data is passed to the LLM-driven script generation phase.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute external tools such as ffmpeg, ffprobe, and a local Python inference script (inference.py) for avatar generation. All calls use list-based arguments, which is a secure practice that prevents shell injection attacks.
[EXTERNAL_DOWNLOADS]: The skill downloads binary model weights (.pth, .safetensors, and .tar files) from GitHub releases (specifically from the OpenTalker, xinntao, and TencentARC organizations) using a PowerShell script (scripts/download_models.ps1). These downloads target well-known open-source repositories associated with the AI models used in the skill.

context-to-video