The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user content by transcribing audio from video files into transcript.json. The agent is then instructed in composition-craft.md to perform a 'role annotation' pass, judging tokens to decide which words to promote or emphasize. This allows potentially malicious instructions spoken in the video to enter the agent's context.
Ingestion points: transcript.json produced by transcribe.cjs.
Boundary markers: Absent; the agent is directed to read and evaluate the raw transcript text.
Capability inventory: The skill can execute shell commands via ffmpeg and node throughout the project lifecycle.
Sanitization: Absent; there is no validation or filtering of the transcript text before the agent analyzes it.
[EXTERNAL_DOWNLOADS]: The skill downloads packages and models during execution.
scripts/transcribe.cjs uses uvx to download and run the whisperx Python package.
scripts/matte.cjs uses the hyperframes CLI to download the u2net_human_seg model (~168 MB) for background removal.
Templates such as modes/standard/_anatomy.md load the gsap library from the jsdelivr CDN.
[COMMAND_EXECUTION]: Most scripts in the scripts/ directory (audio-envelope.cjs, matte.cjs, check-occlusion.cjs, etc.) use child_process modules to execute ffmpeg, ffprobe, and other node scripts to perform video processing and layout validation.
[DYNAMIC_EXECUTION]: Measurement and validation scripts (measure-layout.cjs, check-overflow.cjs, check-rail-climax.cjs) utilize Puppeteer to load local HTML files and execute JavaScript via page.evaluate() to determine the physical bounding boxes of rendered text.

embedded-captions