The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted user-supplied data, including descriptions, image references, and video references, to generate complex video prompts. This constitutes a standard attack surface for AI agents.
Ingestion points: Untrusted data enters via user instructions and reference asset handles like {{Image N}} or {{Video N}} within SKILL.md.
Boundary markers: The skill employs a strict six-section output template (e.g., 【角色/资产锁定】, 【镜头语言总控】, 【事件节拍】) to provide structure and limit accidental obedience to instructions embedded in input data.
Capability inventory: The skill is primarily instructional and does not define access to file-writing tools, network tools, or code execution environments in its configuration.
Sanitization: The instructions enforce precise formatting for asset handles (e.g., {{Image N}}（...参考）) which helps in distinguishing data from instructions.
[COMMAND_EXECUTION]: The skill contains a section on version management that references Git operations such as creating branches, committing changes, and "Syncing Git."
These instructions are intended for maintaining the skill's logic across iterations and are presented as workflow advice.
The agent is directed to perform these actions only upon user request (e.g., "When the user says 'sync Git'").
There is no evidence of high-risk shell command injection or unauthorized system access; the behavior is consistent with a developer assistant persona.
[EXTERNAL_DOWNLOADS]: The skill references a external URL to a well-known service for technical reference.
Evidence: https://x.com/studio_oneroom/status/2075555056985010650 in references/high-density-environment-fight-pattern.md.
This link points to a public tutorial/example on a well-known platform (X/Twitter), which is considered a safe reference for technical documentation.

seedance-fight-director