The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's scaffolding script (new-montage.sh) installs several standard NPM packages including playwright (for browser-based rendering), simple-icons, and various IBM Plex fonts. It also facilitates the download of whisper.cpp transcription models from Hugging Face, which is a well-known and trusted service in the AI community.
[COMMAND_EXECUTION]: The pipeline scripts (ingest.mjs, cut.mjs, compose.mjs, review.sh) execute common CLI utilities such as ffmpeg, ffprobe, whisper-cli, and bc. These tools are essential for media processing and are invoked using safe subprocess methods (execFileSync, spawnSync) that prevent shell injection by avoiding shell parsing of arguments.
[INDIRECT_PROMPT_INJECTION]: As the skill ingests and transcribes external video and audio files, there is a theoretical surface for indirect prompt injection where spoken instructions or embedded text could attempt to influence the agent's behavior during the editing process. However, the skill maintains a human-in-the-loop 'plan gate' where the user must approve the editing plan before final rendering, which serves as a significant security boundary.
Ingestion points: scripts/ingest.mjs (transcribes source video/audio).
Boundary markers: None explicitly defined for data content, but the edit.json control file and docs/plan.md approval step provide process boundaries.
Capability inventory: Subprocess execution of ffmpeg, whisper-cli, node, and browser-based rendering via playwright.
Sanitization: Path sanitization is performed in new-montage.sh, though standard media utilities are responsible for handling binary data safely.

9k-montage