The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of child_process.spawn and spawnSync to execute system-level commands and external tools.
Evidence: scripts/audio.mjs, scripts/lib/tts.mjs, and scripts/lib/bgm.mjs invoke ffmpeg for transcoding, ffprobe for metadata, and npx to run internal CLI tools.
[EXTERNAL_DOWNLOADS]: The skill automatically downloads and installs several third-party Python libraries and fetches remote audio assets at runtime.
Evidence: scripts/lib/bgm.mjs uses pip install to install transformers, torch, soundfile, numpy, and google-genai. scripts/lib/heygen.mjs uses fetch to download audio files from HeyGen's API endpoints.
[DYNAMIC_EXECUTION]: The engine constructs Python scripts as string templates at runtime and executes them using the python3 -c command.
Evidence: scripts/lib/bgm.mjs (in the musicgenScript function) and scripts/lib/tts.mjs (in the ELEVENLABS_PY constant) assemble executable Python code, though they use JSON.stringify to mitigate injection risks into these templates.
[CREDENTIALS_UNSAFE]: The skill logic specifically retrieves and uses sensitive API keys and OAuth tokens from local configuration files and environment variables.
Evidence: scripts/lib/heygen.mjs reads credentials from ~/.heygen/credentials and searches up the directory tree for .env files to load HEYGEN_API_KEY, ELEVENLABS_API_KEY, and GOOGLE_API_KEY.
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and analyze untrusted transcript data to make automated decisions about visual styles, font choices, and animation intensities.
Ingestion points: references/captions/authoring.md and references/captions/motion.md specify reading the full transcript (transcript.json) to detect tone.
Boundary markers: None are explicitly defined to separate instructions from the transcript content during style detection.
Capability inventory: The skill can execute shell commands, write files, and perform network operations.
Sanitization: While basic cleanup is performed in transcript-handling.md, it does not sanitize the narrative content used for style analysis.

hyperframes-media