The Agent Skills Directory

[COMMAND_EXECUTION]: The filmmaking workflow in skills/seedance-audio/SKILL.md instructs the agent to execute a PowerShell script (create_reference_audio.ps1) using the -ExecutionPolicy Bypass flag. This script manages local system processes by attempting to start a Node.js server (server.js) from the user's LOCALAPPDATA directory and transmitting AI-generated JSON payloads to a local API endpoint on localhost:5177 via Invoke-RestMethod.
[COMMAND_EXECUTION]: The repository ships with a variety of Python management scripts, including an installation script (scripts/install_codex_skill.py) and various validation tools (e.g., scripts/validate_skills.py, scripts/content_audit.py). These scripts are designed for automated execution and perform file system operations such as copying repository trees and deleting directories within the user's home folder structure.
[INDIRECT_PROMPT_INJECTION]: The skill's core functionality involves ingesting free-text creative ideas from users to generate complex video prompts and configuration objects (e.g., in skills/seedance-prompt/SKILL.md and skills/seedance-interview/SKILL.md). While the skill utilizes reference tags (e.g., [Image1], @图1) as structural boundaries, the processing of untrusted creative briefs creates a vulnerability surface where adversarial instructions could influence the content of generated prompts or the JSON payloads passed to local execution scripts.

seedance-20