The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection because it ingests user-provided audio, transcribes it, and then interpolates the resulting transcript into a prompt for sub-agents via the PART_BRIEF.md template. A malicious user could potentially embed instructions within the video's audio to influence the sub-agents' behavior during the project building phase.
Ingestion points: The scripts/words.py script saves transcription data from Deepgram into words.json.
Boundary markers: The PART_BRIEF.md template does not currently utilize explicit delimiters or 'ignore' instructions around the transcript block to prevent instructions in the data from being interpreted as commands.
Capability inventory: The skill possesses the ability to write TypeScript/React code and execute rendering commands (npx remotion).
Sanitization: Transcripts are used verbatim without filtering for potential command injection patterns.
[EXTERNAL_DOWNLOADS]: The skill automates the installation of dependencies from trusted and well-known sources. This includes system-level binaries via Homebrew (brew) on macOS, Node.js packages via npm within the rendered projects, the Inter font from Google Fonts, and a headless browser required by the Remotion rendering engine.
[DYNAMIC_EXECUTION]: The skill's primary function involves dynamic execution of code. Sub-agents are tasked with writing TypeScript and React code (src/parts/part-*.tsx) which is then executed by the Remotion engine to generate video frames. This is a legitimate and central feature of the skill's video editing pipeline.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool and Python's subprocess module to coordinate external media tools such as ffmpeg and ffprobe for video analysis, cutting, sound synthesis, and compositing. These operations are conducted using list-based arguments which mitigate common shell injection risks.

cut