The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has a defined surface for indirect prompt injection as it processes untrusted user-supplied media and reference files.
Ingestion points: Data enters the agent context through the intake process described in SKILL.md (Step 1) and media inspection in continuation-handoff.md.
Boundary markers: The skill contains a strong 'Agent Trust Boundary' section in SKILL.md that explicitly instructs the agent to treat text in images, video, and transcripts as source data and to ignore embedded commands.
Capability inventory: The skill includes scripts that execute external binaries like FFmpeg (scripts/extract_last_frame.py) and make network requests (scripts/atlas_seedance_generate.py).
Sanitization: The skill provides behavioral guidelines for the agent to extract only creative facts and ignore operational requests found in data.
[COMMAND_EXECUTION]: Several maintenance and utility scripts use subprocess to execute system commands.
scripts/extract_last_frame.py invokes ffmpeg to process video files. The script includes hardening measures such as protocol whitelisting and timeouts.
scripts/build_masthead_outlines.py uses subprocess to manage virtual environments and install dependencies.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloads and network connections for legitimate operational purposes.
scripts/atlas_seedance_generate.py makes HTTPS requests to api.atlascloud.ai to submit video generation tasks.
scripts/build_masthead_outlines.py performs hash-verified installations of Python packages from PyPI, which is a trusted service.

seedance-20