The Agent Skills Directory

[DYNAMIC_EXECUTION]: The workflow requires the AI agent to generate a custom Python script (video.py) based on transcripts and then execute it via uv run. This dynamic creation and execution of code is a central feature but poses an inherent risk if input data is malicious.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted narration recordings which are transcribed. These transcripts are used to guide the generation of the video.py script, creating an attack surface where narration could theoretically influence the generated code. Ingestion points: Narrated audio files (take.wav) are processed into text transcripts. Boundary markers: No specific delimiters or 'ignore instructions' markers are applied to the transcript content when used as context for script generation. Capability inventory: The skill possesses capabilities to write files, execute Python scripts, and run shell commands (ffmpeg, npx). Sanitization: The skill relies on the LLM to interpret transcriptions as data rather than instructions during the script generation phase.
[EXTERNAL_DOWNLOADS]: The rendering process (render.py) invokes npx to download and execute the hyperframes package. The configuration in templates/hyperframes.json references a registry hosted on a public GitHub repository belonging to HeyGen, a well-known technology provider.
[COMMAND_EXECUTION]: Multiple Python scripts (build.py, master.py, probe.py, render.py) utilize the subprocess module to call system utilities such as ffmpeg and ffprobe for media manipulation, as well as npx for rendering.

narrated-video